diff --git a/bsnes/emulator/emulator.hpp b/bsnes/emulator/emulator.hpp index b720b098a..91e3be523 100644 --- a/bsnes/emulator/emulator.hpp +++ b/bsnes/emulator/emulator.hpp @@ -29,7 +29,7 @@ using namespace nall; namespace Emulator { static const string Name = "bsnes"; - static const string Version = "110.1"; + static const string Version = "110.2"; static const string Author = "byuu"; static const string License = "GPLv3"; static const string Website = "https://byuu.org"; diff --git a/bsnes/target-bsnes/GNUmakefile b/bsnes/target-bsnes/GNUmakefile index 79a3a7050..7c7baa72f 100644 --- a/bsnes/target-bsnes/GNUmakefile +++ b/bsnes/target-bsnes/GNUmakefile @@ -27,9 +27,11 @@ ifeq ($(platform),macos) mkdir -p out/$(name).app/Contents/MacOS/ mkdir -p out/$(name).app/Contents/MacOS/Database/ mkdir -p out/$(name).app/Contents/MacOS/Firmware/ + mkdir -p out/$(name).app/Contents/MacOS/Shaders/ mkdir -p out/$(name).app/Contents/Resources/ mv out/$(name) out/$(name).app/Contents/MacOS/$(name) cp Database/* out/$(name).app/Contents/MacOS/Database/ + cp -r ../shaders/* out/$(name).app/Contents/macOS/Shaders/ cp $(ui)/resource/$(name).plist out/$(name).app/Contents/Info.plist sips -s format icns $(ui)/resource/$(name).png --out out/$(name).app/Contents/Resources/$(name).icns endif @@ -44,6 +46,7 @@ else ifeq ($(platform),macos) mkdir -p ~/Library/Application\ Support/$(name)/ mkdir -p ~/Library/Application\ Support/$(name)/Database/ mkdir -p ~/Library/Application\ Support/$(name)/Firmware/ + mkdir -p ~/Library/Application\ Support/$(name)/Shaders/ cp -R out/$(name).app /Applications/$(name).app else ifneq ($(filter $(platform),linux bsd),) mkdir -p $(prefix)/bin/ @@ -52,12 +55,12 @@ else ifneq ($(filter $(platform),linux bsd),) mkdir -p $(prefix)/share/$(name)/ mkdir -p $(prefix)/share/$(name)/Database/ mkdir -p $(prefix)/share/$(name)/Firmware/ - mkdir -p $(prefix)/share/$(name)/Locale/ + mkdir -p $(prefix)/share/$(name)/Shaders/ cp out/$(name) $(prefix)/bin/$(name) cp $(ui)/resource/$(name).desktop $(prefix)/share/applications/$(name).desktop cp $(ui)/resource/$(name).png $(prefix)/share/icons/$(name).png cp Database/* $(prefix)/share/$(name)/Database/ - cp Locale/* $(prefix)/share/$(name)/Locale/ + cp -r ../shaders/* $(prefix)/share/$(name)/Shaders/ endif uninstall: diff --git a/bsnes/target-bsnes/presentation/presentation.cpp b/bsnes/target-bsnes/presentation/presentation.cpp index 94b8e2046..700454e49 100644 --- a/bsnes/target-bsnes/presentation/presentation.cpp +++ b/bsnes/target-bsnes/presentation/presentation.cpp @@ -530,7 +530,7 @@ auto Presentation::updateShaders() -> void { }); shaders.append(blur); - auto location = locate("shaders/"); + auto location = locate("Shaders/"); if(settings.video.driver == "OpenGL 3.2") { for(auto shader : directory::folders(location, "*.shader")) { diff --git a/bsnes/target-libretro/GNUmakefile b/bsnes/target-libretro/GNUmakefile index 4c4cf7fb3..104128e3d 100644 --- a/bsnes/target-libretro/GNUmakefile +++ b/bsnes/target-libretro/GNUmakefile @@ -1,5 +1,7 @@ name := libretro.so -flags += -Wno-narrowing -Wno-multichar -fopenmp -g -fPIC +local := false +openmp := true +flags += -Wno-narrowing -Wno-multichar -g -fPIC objects := libretro $(objects) objects := $(patsubst %,obj/%.o,$(objects)) @@ -13,4 +15,14 @@ else ifeq ($(platform),windows) $(strip $(compiler) -o out/bsnes_libretro.dll -shared $(objects) -Wl,--no-undefined -Wl,--version-script=target-libretro/link.T -static-libgcc -static-libstdc++ -Wl,-Bstatic -lws2_32 -lpthread -lgomp -Wl,-Bdynamic) else ifeq ($(platform),macos) $(strip $(compiler) -o out/bsnes_libretro.dylib -shared $(objects) -lpthread -ldl) +else ifeq ($(platform), ios-arm64) + ifeq ($(IOSSDK),) + IOSSDK := $(shell xcodebuild -version -sdk iphoneos Path) + endif + $(strip c++ -arch arm64 -marm -miphoneos-version-min=11.0 -isysroot $(IOSSDK) -o out/bsnes_libretro_ios.dylib -shared $(objects) -lpthread -ldl) +else ifeq ($(platform), tvos-arm64) + ifeq ($(IOSSDK),) + IOSSDK := $(shell xcodebuild -version -sdk appletvos Path) + endif + $(strip c++ -arch arm64 -marm -mtvos-version-min=11.0 -isysroot $(IOSSDK) -o out/bsnes_libretro_tvos.dylib -shared $(objects) -lpthread -ldl) endif diff --git a/bsnes/target-libretro/program.cpp b/bsnes/target-libretro/program.cpp index c07365c72..cbfa4c374 100644 --- a/bsnes/target-libretro/program.cpp +++ b/bsnes/target-libretro/program.cpp @@ -146,6 +146,9 @@ auto Program::load() -> void { //fixes an errant scanline on the title screen due to writing to PPU registers too late if(title == "ADVENTURES OF FRANKEN" && region == "PAL") emulator->configure("Hacks/PPU/RenderCycle", 32); + //fixes an errant scanline on the title screen due to writing to PPU registers too late + if(title == "FIREPOWER 2000") emulator->configure("Hacks/PPU/RenderCycle", 32); + emulator->power(); } @@ -288,7 +291,7 @@ auto Program::openRomSuperFamicom(string name, vfs::file::mode mode) -> shared_p string save_path; auto suffix = Location::suffix(base_name); - auto base = Location::base(base_name); + auto base = Location::base(base_name.transform("\\", "/")); const char *save = nullptr; if (environ_cb && environ_cb(RETRO_ENVIRONMENT_GET_SAVE_DIRECTORY, &save) && save) diff --git a/bsnes/target-libretro/resources.hpp b/bsnes/target-libretro/resources.hpp index b0ba3fa6a..1a4386263 100644 --- a/bsnes/target-libretro/resources.hpp +++ b/bsnes/target-libretro/resources.hpp @@ -1,4 +1,4 @@ -const unsigned char boardsbml[30846] = { +const unsigned char boardsbml[31025] = { 100,97,116,97,98,97,115,101,10,32,32,114,101,118,105,115,105,111,110,58,32,50,48,49,56,45,48,55,45,50,53,10, 10,47,47,66,111,97,114,100,115,32,40,80,114,111,100,117,99,116,105,111,110,41,10,10,100,97,116,97,98,97,115,101, 10,32,32,114,101,118,105,115,105,111,110,58,32,50,48,49,56,45,48,53,45,49,54,10,10,98,111,97,114,100,58,32, @@ -571,398 +571,404 @@ const unsigned char boardsbml[30846] = { 109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32, 32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102, 102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58, - 48,48,48,48,45,102,102,102,102,10,10,47,47,66,111,97,114,100,115,32,40,71,101,110,101,114,105,99,41,10,10,100, - 97,116,97,98,97,115,101,10,32,32,114,101,118,105,115,105,111,110,58,32,50,48,49,56,45,48,55,45,50,53,10,10, - 98,111,97,114,100,58,32,65,82,77,45,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121, - 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97, - 100,100,114,101,115,115,61,48,48,45,55,100,44,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115, - 107,61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,54,102,44,99, - 48,45,101,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109, - 111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97, - 112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10, - 32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,65,82,77,54,10,32,32, - 32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,51,56,48,48,45,51, - 56,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61, - 80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,65,82,77,54,10,32,32,32,32,109,101, - 109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105, - 116,101,99,116,117,114,101,61,65,82,77,54,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77, + 48,48,48,48,45,102,102,102,102,10,10,47,47,66,111,97,114,100,115,32,40,80,114,111,116,111,116,121,112,101,115,41, + 10,10,98,111,97,114,100,58,32,83,72,86,67,45,52,80,86,53,66,45,48,49,10,32,32,109,101,109,111,114,121,32, + 116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112, + 32,97,100,100,114,101,115,115,61,48,48,45,55,100,44,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109, + 97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100, + 44,99,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,10,47,47, + 66,111,97,114,100,115,32,40,71,101,110,101,114,105,99,41,10,10,100,97,116,97,98,97,115,101,10,32,32,114,101,118, + 105,115,105,111,110,58,32,50,48,49,56,45,48,55,45,50,53,10,10,98,111,97,114,100,58,32,65,82,77,45,76,79, + 82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, + 116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,44, + 56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32, + 109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,54,102,44,99,48,45,101,102,58,48,48,48,48,45,55,102,102, + 102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32, + 99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45, + 55,100,44,102,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,112,114,111,99,101,115,115,111,114,32,97, + 114,99,104,105,116,101,99,116,117,114,101,61,65,82,77,54,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, + 61,48,48,45,51,102,44,56,48,45,98,102,58,51,56,48,48,45,51,56,102,102,10,32,32,32,32,109,101,109,111,114, + 121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105, + 116,101,99,116,117,114,101,61,65,82,77,54,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77, 32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,65,82,77,54,10, - 32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,66,83,45,72,73,82,79,77,45, - 82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114, - 111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57, - 102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,53, - 102,44,99,48,45,100,102,58,48,48,48,48,45,102,102,102,102,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61, - 82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48, - 48,48,10,32,32,115,108,111,116,32,116,121,112,101,61,66,83,77,101,109,111,114,121,10,32,32,32,32,109,97,112,32, - 97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32, - 32,32,109,97,112,32,97,100,100,114,101,115,115,61,54,48,45,55,100,44,101,48,45,102,102,58,48,48,48,48,45,102, - 102,102,102,10,10,98,111,97,114,100,58,32,66,83,45,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114, - 121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61, - 48,120,48,48,48,48,48,48,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100, - 114,101,115,115,61,50,48,45,51,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,49,48,48,48, - 48,48,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56, - 48,45,57,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,50,48,48,48,48,48,32,109,97,115, - 107,61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,97,48,45,98,102,58,56, - 48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,49,48,48,48,48,48,32,109,97,115,107,61,48,120,56,48, - 48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118, - 101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48, - 48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,115,108,111,116,32,116,121,112,101,61, - 66,83,77,101,109,111,114,121,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,101,102,58,48, - 48,48,48,45,102,102,102,102,10,10,98,111,97,114,100,58,32,66,83,45,77,67,67,45,82,65,77,10,32,32,109,101, + 32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97, + 32,97,114,99,104,105,116,101,99,116,117,114,101,61,65,82,77,54,10,32,32,32,32,111,115,99,105,108,108,97,116,111, + 114,10,10,98,111,97,114,100,58,32,66,83,45,72,73,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32, + 116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112, + 32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58,56,48,48,48,45,102,102,102,102,10,32, + 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,53,102,44,99,48,45,100,102,58,48,48,48,48,45, + 102,102,102,102,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83, + 97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58, + 54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,115,108,111,116,32,116,121,112, + 101,61,66,83,77,101,109,111,114,121,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102, + 44,97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, + 61,54,48,45,55,100,44,101,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,10,98,111,97,114,100,58,32,66, + 83,45,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111, + 110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48, + 45,49,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,48,48,48,48,48,48,32,109,97,115,107, + 61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,58,56,48, + 48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,49,48,48,48,48,48,32,109,97,115,107,61,48,120,56,48,48, + 48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48,45,57,102,58,56,48,48,48,45,102,102,102, + 102,32,98,97,115,101,61,48,120,50,48,48,48,48,48,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32, + 109,97,112,32,97,100,100,114,101,115,115,61,97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101, + 61,48,120,49,48,48,48,48,48,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109,111,114,121,32,116, + 121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100, + 114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61, + 48,120,56,48,48,48,10,32,32,115,108,111,116,32,116,121,112,101,61,66,83,77,101,109,111,114,121,10,32,32,32,32, + 109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,101,102,58,48,48,48,48,45,102,102,102,102,10,10,98,111,97, + 114,100,58,32,66,83,45,77,67,67,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77, + 32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,49,48, + 45,49,55,58,53,48,48,48,45,53,102,102,102,32,109,97,115,107,61,48,120,102,48,48,48,10,32,32,112,114,111,99, + 101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,77,67,67,10,32,32,32,32,109,97,112,32,97,100,100, + 114,101,115,115,61,48,48,45,48,102,58,53,48,48,48,45,53,102,102,102,10,32,32,32,32,109,99,117,10,32,32,32, + 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45, + 102,102,102,102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45, + 102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50, + 48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102,10,32,32,32,32,32,32,109,101,109,111,114, + 121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32, + 32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,111,119,110,108,111,97, + 100,10,32,32,32,32,32,32,115,108,111,116,32,116,121,112,101,61,66,83,77,101,109,111,114,121,10,10,98,111,97,114, + 100,58,32,66,83,45,83,65,49,45,82,65,77,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116, + 101,99,116,117,114,101,61,87,54,53,67,56,49,54,83,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61, + 48,48,45,51,102,44,56,48,45,98,102,58,50,50,48,48,45,50,51,102,102,10,32,32,32,32,109,99,117,10,32,32, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48, + 45,102,102,102,102,32,109,97,115,107,61,48,120,52,48,56,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,101,109,111, + 114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32, + 32,32,115,108,111,116,32,116,121,112,101,61,66,83,77,101,109,111,114,121,10,32,32,32,32,109,101,109,111,114,121,32, + 116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32, + 97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,115,105, + 122,101,61,48,120,50,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,52, + 102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32, + 99,111,110,116,101,110,116,61,73,110,116,101,114,110,97,108,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101, + 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,51,48,48,48,45,51,55,102,102,32,115,105,122,101,61,48,120, + 56,48,48,10,10,98,111,97,114,100,58,32,69,86,69,78,84,45,67,67,57,50,10,32,32,109,101,109,111,114,121,32, + 116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107, + 61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,109,97,110,117,102,97,99,116,117,114,101,114, + 61,78,69,67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,56,50,49,52,10,32,32,32,32,105, + 100,101,110,116,105,102,105,101,114,58,32,67,97,109,112,117,115,32,67,104,97,108,108,101,110,103,101,32,39,57,50,10, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,44,101,48,58,48,48,48,48,10,32,32,32,32,109, + 99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102, + 58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77, + 32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121, + 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101,108,45,49,10,32,32,32,32,32,32,109,101,109, + 111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101,108,45,50,10,32,32,32, + 32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101,108, + 45,51,10,32,32,32,32,100,105,112,10,32,32,112,114,111,99,101,115,115,111,114,32,109,97,110,117,102,97,99,116,117, + 114,101,114,61,78,69,67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32, + 32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,56,48,48,48,45,102,102, + 102,102,32,109,97,115,107,61,48,120,55,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82, + 79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61, + 117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116, + 101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32, + 32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97, + 114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116, + 111,114,10,10,98,111,97,114,100,58,32,69,86,69,78,84,45,80,70,57,52,10,32,32,109,101,109,111,114,121,32,116, + 121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100, + 114,101,115,115,61,51,48,45,51,102,44,98,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61, + 48,120,101,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,109,97,110,117,102,97,99,116,117,114,101,114,61, + 78,69,67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,56,50,49,52,10,32,32,32,32,105,100, + 101,110,116,105,102,105,101,114,58,32,80,111,119,101,114,70,101,115,116,32,39,57,52,10,32,32,32,32,109,97,112,32, + 97,100,100,114,101,115,115,61,49,48,44,50,48,58,54,48,48,48,10,32,32,32,32,109,99,117,10,32,32,32,32,32, + 32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102, + 102,102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45, + 102,102,102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101, + 110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77, + 32,99,111,110,116,101,110,116,61,76,101,118,101,108,45,49,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121, + 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101,108,45,50,10,32,32,32,32,32,32,109,101,109, + 111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101,108,45,51,10,32,32,32, + 32,100,105,112,10,32,32,112,114,111,99,101,115,115,111,114,32,109,97,110,117,102,97,99,116,117,114,101,114,61,78,69, + 67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97, + 100,100,114,101,115,115,61,48,48,45,48,102,44,56,48,45,56,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115, + 107,61,48,120,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116, + 101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53, + 10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116, + 97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114, + 121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99, + 116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97, + 114,100,58,32,69,88,72,73,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111, + 110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48, + 45,51,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32, + 109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,58,48,48,48,48,45,102,102,102,102,32,98,97,115,101, + 61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48,45,98,102,58, + 56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,109,97,112,32, + 97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99, + 48,48,48,48,48,10,10,98,111,97,114,100,58,32,69,88,72,73,82,79,77,45,82,65,77,10,32,32,109,101,109,111, + 114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32, + 109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101, + 61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,58, + 48,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32, + 97,100,100,114,101,115,115,61,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99, + 48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48, + 45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112, + 101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101, + 115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120, + 101,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,58,48,48,48,48,45, + 55,102,102,102,10,10,98,111,97,114,100,58,32,69,88,72,73,82,79,77,45,82,65,77,45,83,72,65,82,80,82,84, + 67,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103, + 114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,58,56,48,48,48,45,102, + 102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115, + 115,61,52,48,45,55,100,58,48,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32, + 109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48, + 45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,109,101, 109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,49,48,45,49,55,58,53,48,48,48,45,53,102,102,102,32,109,97,115,107,61, - 48,120,102,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,77,67, - 67,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,48,102,58,53,48,48,48,45,53,102,102, - 102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51, - 102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114, - 101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32, - 109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102, - 102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61, - 80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111, - 110,116,101,110,116,61,68,111,119,110,108,111,97,100,10,32,32,32,32,32,32,115,108,111,116,32,116,121,112,101,61,66, - 83,77,101,109,111,114,121,10,10,98,111,97,114,100,58,32,66,83,45,83,65,49,45,82,65,77,10,32,32,112,114,111, - 99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,87,54,53,67,56,49,54,83,10,32,32,32, - 32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,50,50,48,48,45,50,51, - 102,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45, - 51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,52,48,56,48,48,48, - 10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102, - 102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116, - 61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,115,108,111,116,32,116,121,112,101,61,66,83,77,101,109,111,114, - 121,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97, - 118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102, - 58,54,48,48,48,45,55,102,102,102,32,115,105,122,101,61,48,120,50,48,48,48,10,32,32,32,32,32,32,109,97,112, - 32,97,100,100,114,101,115,115,61,52,48,45,52,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,101,109, - 111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,73,110,116,101,114,110,97,108,10,32,32, - 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,51,48,48,48, - 45,51,55,102,102,32,115,105,122,101,61,48,120,56,48,48,10,10,98,111,97,114,100,58,32,69,86,69,78,84,45,67, - 67,57,50,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97, - 118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48, - 48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114, - 32,109,97,110,117,102,97,99,116,117,114,101,114,61,78,69,67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117, - 80,68,55,56,50,49,52,10,32,32,32,32,105,100,101,110,116,105,102,105,101,114,58,32,67,97,109,112,117,115,32,67, - 104,97,108,108,101,110,103,101,32,39,57,50,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,44, - 101,48,58,48,48,48,48,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115, - 115,61,48,48,45,49,102,44,56,48,45,57,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,101, + 97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102, + 32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45, + 55,100,58,48,48,48,48,45,55,102,102,102,10,32,32,114,116,99,32,109,97,110,117,102,97,99,116,117,114,101,114,61, + 83,104,97,114,112,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98, + 102,58,50,56,48,48,45,50,56,48,49,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,84,67,32, + 99,111,110,116,101,110,116,61,84,105,109,101,32,109,97,110,117,102,97,99,116,117,114,101,114,61,83,104,97,114,112,10, + 10,98,111,97,114,100,58,32,69,88,76,79,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79, + 77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115, + 115,61,48,48,45,55,100,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,56,48,48,48,32, + 98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48, + 45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,56,48,48,48,32,98,97,115,101, + 61,48,120,48,48,48,48,48,48,10,10,98,111,97,114,100,58,32,69,88,76,79,82,79,77,45,82,65,77,10,32,32, + 109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,58,56,48,48,48,45,102,102,102,102,32, + 109,97,115,107,61,48,120,56,48,56,48,48,48,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32, + 109,97,112,32,97,100,100,114,101,115,115,61,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107, + 61,48,120,56,48,56,48,48,48,32,98,97,115,101,61,48,120,48,48,48,48,48,48,10,32,32,109,101,109,111,114,121, + 32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97, + 100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115, + 107,61,48,120,56,48,48,48,10,10,98,111,97,114,100,58,32,69,88,78,69,67,45,76,79,82,79,77,10,32,32,109, + 101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32, + 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,44,56,48,45,102,102,58,56,48,48,48,45, + 102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99, + 104,105,116,101,99,116,117,114,101,61,117,80,68,57,54,48,53,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101, + 115,115,61,54,48,45,54,55,44,101,48,45,101,55,58,48,48,48,48,45,51,102,102,102,10,32,32,32,32,109,101,109, + 111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99, + 104,105,116,101,99,116,117,114,101,61,117,80,68,57,54,48,53,48,10,32,32,32,32,109,101,109,111,114,121,32,116,121, + 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101, + 61,117,80,68,57,54,48,53,48,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111, + 110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,57,54,48,53,48, + 10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,54,56,45,54,102,44,101,56,45,101,102,58,48, + 48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,111,115,99,105,108,108,97, + 116,111,114,10,10,98,111,97,114,100,58,32,69,88,83,80,67,55,49,49,48,45,82,65,77,45,69,80,83,79,78,82, + 84,67,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,69,120,112, + 97,110,115,105,111,110,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,52,102,58,48,48,48, + 48,45,102,102,102,102,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,83,80, + 67,55,49,49,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98, + 102,58,52,56,48,48,45,52,56,51,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,53,48,44,53, + 56,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107, + 61,48,120,56,48,48,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102, + 102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,32,32, + 109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10, + 32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97, + 116,97,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83, + 97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98, + 102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,114,116,99,32,109,97, + 110,117,102,97,99,116,117,114,101,114,61,69,112,115,111,110,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, + 61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,52,48,45,52,56,52,50,10,32,32,32,32,109,101,109,111,114, + 121,32,116,121,112,101,61,82,84,67,32,99,111,110,116,101,110,116,61,84,105,109,101,32,109,97,110,117,102,97,99,116, + 117,114,101,114,61,69,112,115,111,110,10,10,98,111,97,114,100,58,32,71,66,45,76,79,82,79,77,10,32,32,109,101, 109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32, - 32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101, - 108,45,49,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, - 116,61,76,101,118,101,108,45,50,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32, - 99,111,110,116,101,110,116,61,76,101,118,101,108,45,51,10,32,32,32,32,100,105,112,10,32,32,112,114,111,99,101,115, - 115,111,114,32,109,97,110,117,102,97,99,116,117,114,101,114,61,78,69,67,32,97,114,99,104,105,116,101,99,116,117,114, - 101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44, - 97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,55,102,102,102,10,32,32,32,32, - 109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32, - 97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32, - 116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117, - 114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99, - 111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53, - 10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,69,86,69,78,84,45,80,70, - 57,52,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118, - 101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,51,48,45,51,102,44,98,48,45,98,102,58,54,48, - 48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32, - 109,97,110,117,102,97,99,116,117,114,101,114,61,78,69,67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80, - 68,55,56,50,49,52,10,32,32,32,32,105,100,101,110,116,105,102,105,101,114,58,32,80,111,119,101,114,70,101,115,116, - 32,39,57,52,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,49,48,44,50,48,58,54,48,48,48,10, - 32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44, - 56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115, - 115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116, - 121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101, - 109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101,108,45,49,10,32,32, - 32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,76,101,118,101, - 108,45,50,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, - 116,61,76,101,118,101,108,45,51,10,32,32,32,32,100,105,112,10,32,32,112,114,111,99,101,115,115,111,114,32,109,97, - 110,117,102,97,99,116,117,114,101,114,61,78,69,67,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55, - 55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,48,102,44,56,48,45,56,102,58, + 32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,44,56,48,45,102,102,58,56,48,48,48,45,102, + 102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61, + 52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48, + 48,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,73,67,68,32,114,101,118, + 105,115,105,111,110,61,50,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48, + 45,98,102,58,54,48,48,48,45,54,55,102,102,44,55,48,48,48,45,55,102,102,102,10,32,32,32,32,109,101,109,111, + 114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,66,111,111,116,32,97,114,99,104,105,116,101, + 99,116,117,114,101,61,76,82,51,53,57,48,50,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,32,32,32, + 32,115,108,111,116,32,116,121,112,101,61,71,97,109,101,66,111,121,10,10,98,111,97,114,100,58,32,71,83,85,45,82, + 65,77,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,71,83,85,10, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,51,48,48,48, + 45,51,52,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, + 116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51, + 102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,53,102,44,99,48,45,100,102,58,48,48,48,48, + 45,102,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110, + 116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56, + 48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,115,105,122,101,61,48,120,50,48,48,48,10,32,32,32,32,32, + 32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,49,44,102,48,45,102,49,58,48,48,48,48,45,102,102, + 102,102,10,10,98,111,97,114,100,58,32,72,73,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82, + 79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101, + 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,109,97,112, + 32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,10, + 98,111,97,114,100,58,32,72,73,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82, + 79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101, + 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,109,97,112, + 32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32, + 32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32, + 32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55, + 102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,10,98,111,97,114,100,58,32,72,73,84,65,67,72,73,45, + 76,79,82,79,77,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,72, + 71,53,49,66,83,49,54,57,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56, + 48,45,98,102,58,54,99,48,48,45,54,102,102,102,44,55,99,48,48,45,55,102,102,102,10,32,32,32,32,109,101,109, + 111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32, + 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45, + 102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101, + 61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114, + 101,115,115,61,55,48,45,55,55,58,48,48,48,48,45,55,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116, + 121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114, + 101,61,72,71,53,49,66,83,49,54,57,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32, + 99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,72,71,53,49,66,83, + 49,54,57,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98, + 102,58,54,48,48,48,45,54,98,102,102,44,55,48,48,48,45,55,98,102,102,32,109,97,115,107,61,48,120,102,48,48, + 48,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,76,79,82,79,77,10,32, + 32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109, + 10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,44,56,48,45,102,102,58,56,48,48, + 48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,10,98,111,97,114,100,58,32,76,79,82,79,77, + 45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80, + 114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,44,56,48,45, + 102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109,111,114, + 121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32, + 97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32,109,97, + 115,107,61,48,120,56,48,48,48,10,10,98,111,97,114,100,58,32,76,79,82,79,77,45,82,65,77,35,65,10,32,32, + 109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48, + 45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61, + 82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, + 61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48, + 48,48,10,10,98,111,97,114,100,58,32,78,69,67,45,72,73,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121, + 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97, + 100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32, + 32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,102,102, + 102,102,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55, + 55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58, 54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32, 116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101, 99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79, 77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55, 55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61, 68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,111,115, - 99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,69,88,72,73,82,79,77,10,32,32,109,101,109,111,114, - 121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61, - 48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,58,48, - 48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97, - 100,100,114,101,115,115,61,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48, - 48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45, - 102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,10,98,111,97,114,100,58,32,69,88,72,73,82, - 79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116, - 61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,58,56, - 48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97, - 100,100,114,101,115,115,61,52,48,45,55,100,58,48,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48, - 48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48,45,98,102,58,56,48,48,48,45, - 102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101, - 115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48, - 10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10, - 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48, - 45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115, - 115,61,55,48,45,55,100,58,48,48,48,48,45,55,102,102,102,10,10,98,111,97,114,100,58,32,69,88,72,73,82,79, - 77,45,82,65,77,45,83,72,65,82,80,82,84,67,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77, - 32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,48,48,45,51,102,58,56,48,48,48,45,102,102,102,102,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32, - 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,58,48,48,48,48,45,102,102,102,102,32,98, - 97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48,45, - 98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61, - 48,120,99,48,48,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101, - 110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48, - 45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,32,32,109,97, - 112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,58,48,48,48,48,45,55,102,102,102,10,32,32,114,116,99,32, - 109,97,110,117,102,97,99,116,117,114,101,114,61,83,104,97,114,112,10,32,32,32,32,109,97,112,32,97,100,100,114,101, - 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,50,56,48,48,45,50,56,48,49,10,32,32,32,32,109,101,109, - 111,114,121,32,116,121,112,101,61,82,84,67,32,99,111,110,116,101,110,116,61,84,105,109,101,32,109,97,110,117,102,97, - 99,116,117,114,101,114,61,83,104,97,114,112,10,10,98,111,97,114,100,58,32,69,88,76,79,82,79,77,10,32,32,109, - 101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32, - 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100,58,56,48,48,48,45,102,102,102,102,32,109, - 97,115,107,61,48,120,56,48,56,48,48,48,32,98,97,115,101,61,48,120,52,48,48,48,48,48,10,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61, - 48,120,56,48,56,48,48,48,32,98,97,115,101,61,48,120,48,48,48,48,48,48,10,10,98,111,97,114,100,58,32,69, - 88,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110, - 116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45, - 55,100,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,56,48,48,48,32,98,97,115,101,61, - 48,120,52,48,48,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,56,48,45,102,102,58,56, - 48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,56,48,48,48,32,98,97,115,101,61,48,120,48,48, - 48,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83, - 97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58, - 48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,10,98,111,97,114,100,58,32,69,88, - 78,69,67,45,76,79,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116, - 101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55, - 100,44,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32, - 112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,57,54,48,53,48,10, - 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,54,48,45,54,55,44,101,48,45,101,55,58,48,48,48,48, - 45,51,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, - 116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,57,54,48,53,48,10, - 32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97, - 32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,57,54,48,53,48,10,32,32,32,32,109,101,109,111,114, - 121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99, - 116,117,114,101,61,117,80,68,57,54,48,53,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61, - 54,56,45,54,102,44,101,56,45,101,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48, - 48,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,69,88,83,80,67,55,49, - 49,48,45,82,65,77,45,69,80,83,79,78,82,84,67,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79, - 77,32,99,111,110,116,101,110,116,61,69,120,112,97,110,115,105,111,110,10,32,32,32,32,109,97,112,32,97,100,100,114, - 101,115,115,61,52,48,45,52,102,58,48,48,48,48,45,102,102,102,102,10,32,32,112,114,111,99,101,115,115,111,114,32, - 105,100,101,110,116,105,102,105,101,114,61,83,80,67,55,49,49,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101, - 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,48,48,45,52,56,51,102,10,32,32,32,32,109,97,112, - 32,97,100,100,114,101,115,115,61,53,48,44,53,56,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,99,117, - 10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56, - 48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,48,48,10,32,32,32,32,32,32,109,97,112, - 32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120, - 99,48,48,48,48,48,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110, - 116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82, - 79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61, - 82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101, - 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120, - 101,48,48,48,10,32,32,114,116,99,32,109,97,110,117,102,97,99,116,117,114,101,114,61,69,112,115,111,110,10,32,32, - 32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,52,48,45,52, - 56,52,50,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,84,67,32,99,111,110,116,101,110,116,61, - 84,105,109,101,32,109,97,110,117,102,97,99,116,117,114,101,114,61,69,112,115,111,110,10,10,98,111,97,114,100,58,32, - 71,66,45,76,79,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101, - 110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,55,100, - 44,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32, - 32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,55,102, - 102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116, - 105,102,105,101,114,61,73,67,68,32,114,101,118,105,115,105,111,110,61,50,10,32,32,32,32,109,97,112,32,97,100,100, - 114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,54,55,102,102,44,55,48,48,48,45, - 55,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116, - 61,66,111,111,116,32,97,114,99,104,105,116,101,99,116,117,114,101,61,76,82,51,53,57,48,50,10,32,32,32,32,111, - 115,99,105,108,108,97,116,111,114,10,32,32,32,32,115,108,111,116,32,116,121,112,101,61,71,97,109,101,66,111,121,10, - 10,98,111,97,114,100,58,32,71,83,85,45,82,65,77,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104, - 105,116,101,99,116,117,114,101,61,71,83,85,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45, - 51,102,44,56,48,45,98,102,58,51,48,48,48,45,51,52,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121, - 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,97,112, - 32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109, - 97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45, - 53,102,44,99,48,45,100,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121, - 112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100, - 100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,115,105,122,101, - 61,48,120,50,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,49,44, - 102,48,45,102,49,58,48,48,48,48,45,102,102,102,102,10,10,98,111,97,114,100,58,32,72,73,82,79,77,10,32,32, - 109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10, - 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48, - 45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102, - 102,58,48,48,48,48,45,102,102,102,102,10,10,98,111,97,114,100,58,32,72,73,82,79,77,45,82,65,77,10,32,32, - 109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10, - 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48, - 45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45,102, - 102,58,48,48,48,48,45,102,102,102,102,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111, - 110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102, - 44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,10,98,111, - 97,114,100,58,32,72,73,84,65,67,72,73,45,76,79,82,79,77,10,32,32,112,114,111,99,101,115,115,111,114,32,97, - 114,99,104,105,116,101,99,116,117,114,101,61,72,71,53,49,66,83,49,54,57,10,32,32,32,32,109,97,112,32,97,100, - 100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,99,48,48,45,54,102,102,102,44,55,99,48,48, - 45,55,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, - 116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51, - 102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32, - 32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32, - 32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,55,58,48,48,48,48,45,55,102,102,102, - 10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116, - 97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,72,71,53,49,66,83,49,54,57,10,32,32,32,32,109,101,109, - 111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116, - 101,99,116,117,114,101,61,72,71,53,49,66,83,49,54,57,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101, - 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,54,98,102,102,44,55,48,48,48,45,55,98, - 102,102,32,109,97,115,107,61,48,120,102,48,48,48,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98, - 111,97,114,100,58,32,76,79,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111, - 110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48, - 45,55,100,44,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10, - 10,98,111,97,114,100,58,32,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61, - 82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114, - 101,115,115,61,48,48,45,55,100,44,56,48,45,102,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48, - 120,56,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61, - 83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102, - 58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,10,98,111,97,114,100,58,32,76, - 79,82,79,77,45,82,65,77,35,65,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110, - 116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45, - 51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32, - 32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32, - 32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,102, - 102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,10,98,111,97,114,100,58,32,78,69,67,45,72,73,82,79, - 77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103, - 114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58, - 56,48,48,48,45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44, - 99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104, - 105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,48,48,45,49,102,44,56,48,45,57,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,102,102, - 102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114, - 111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109, - 101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104, - 105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101, - 61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117, - 80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,78,69, - 67,45,72,73,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111, - 110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48, - 45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114, - 101,115,115,61,52,48,45,55,100,44,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,109,101,109,111, - 114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112, - 32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109, - 97,115,107,61,48,120,101,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116, - 117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49, - 102,44,56,48,45,57,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,102,102,102,10,32,32,32, + 99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,78,69,67,45,72,73,82,79,77,45,82,65,77,10,32, + 32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109, + 10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48, + 48,45,102,102,102,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,55,100,44,99,48,45, + 102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99, + 111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51, + 102,44,97,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32, + 112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32, + 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58,54,48,48,48,45, + 55,102,102,102,32,109,97,115,107,61,48,120,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61, + 82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101, + 61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110, + 116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32, + 32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32, + 97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97, + 116,111,114,10,10,98,111,97,114,100,58,32,78,69,67,45,76,79,82,79,77,10,32,32,109,101,109,111,114,121,32,116, + 121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32, + 97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58,56,48,48,48,45,102,102,102,102,32,109,97, + 115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117, + 114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,51,48,45,51,102, + 44,98,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,51,102,102,102,10,32,32,32, 32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109, 32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121, 32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116, 117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32, 99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50, 53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,78,69,67,45,76,79,82, - 79,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111, - 103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102, - 58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115, - 111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32, - 97,100,100,114,101,115,115,61,51,48,45,51,102,44,98,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97, - 115,107,61,48,120,51,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111, - 110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55, - 50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68, - 97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109, - 111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116, - 101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98, - 111,97,114,100,58,32,78,69,67,45,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112, - 101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100, - 100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107, - 61,48,120,56,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110, - 116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45, - 102,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101, - 115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97, - 112,32,97,100,100,114,101,115,115,61,54,48,45,54,102,44,101,48,45,101,102,58,48,48,48,48,45,55,102,102,102,32, - 109,97,115,107,61,48,120,51,102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32, - 99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68, - 55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116, - 61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109, - 101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104, - 105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10, - 10,98,111,97,114,100,58,32,78,69,67,45,76,79,82,79,77,45,82,65,77,35,65,10,32,32,109,101,109,111,114,121, - 32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97, - 112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58,56,48,48,48,45,102,102,102,102,32, - 109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111, - 110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100, - 44,102,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99, - 104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115, - 115,61,50,48,45,51,102,44,97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,51, - 102,102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61, - 80,114,111,103,114,97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32, - 32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114, - 99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121, - 112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101, - 61,117,80,68,55,55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32, - 79,66,67,49,45,76,79,82,79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77, - 32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48, - 48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,79,66,67,49,10,32, - 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45, - 55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,55,48,45,55,49,44,102,48,45,102,49,58,54,48,48,48,45,55,102,102,102,44,101,48,48,48,45,102,102,102,102, - 32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77, - 32,99,111,110,116,101,110,116,61,83,97,118,101,10,10,98,111,97,114,100,58,32,83,65,49,45,82,65,77,10,32,32, - 112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,87,54,53,67,56,49,54,83,10, - 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,50,50,48,48, - 45,50,51,102,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61, - 48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,52,48,56, - 48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48, - 45,102,102,102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116, - 101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32, - 99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48, - 48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,115,105,122,101,61,48,120,50,48,48,48, - 10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48,45,52,102,58,48,48,48,48,45,102,102, - 102,102,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,73, - 110,116,101,114,110,97,108,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44, - 56,48,45,98,102,58,51,48,48,48,45,51,55,102,102,32,115,105,122,101,61,48,120,56,48,48,10,10,98,111,97,114, - 100,58,32,83,68,68,49,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,83, - 68,68,49,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58, - 52,56,48,48,45,52,56,48,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114, - 101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32, - 109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32, - 32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97, - 109,10,10,98,111,97,114,100,58,32,83,68,68,49,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101, - 61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115, + 79,77,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116, + 61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56, + 48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,109,101,109, + 111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,109,97, + 112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,55,102,102,102,32, + 109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99, + 116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,54,48,45, + 54,102,44,101,48,45,101,102,58,48,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,51,102,102,102,10,32, + 32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114, + 97,109,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111, + 114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101, + 99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65, + 77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55, + 55,50,53,10,32,32,32,32,111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,78,69,67,45,76, + 79,82,79,77,45,82,65,77,35,65,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110, + 116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45, + 49,102,44,56,48,45,57,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32, + 32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32, + 32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,102, + 102,102,102,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68, + 55,55,50,53,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102, + 58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,51,102,102,102,10,32,32,32,32,109,101,109,111,114, + 121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,32,97,114,99,104,105, + 116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61, + 82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80, + 68,55,55,50,53,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110, + 116,61,68,97,116,97,32,97,114,99,104,105,116,101,99,116,117,114,101,61,117,80,68,55,55,50,53,10,32,32,32,32, + 111,115,99,105,108,108,97,116,111,114,10,10,98,111,97,114,100,58,32,79,66,67,49,45,76,79,82,79,77,45,82,65, + 77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103, + 114,97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58, + 56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111, + 114,32,105,100,101,110,116,105,102,105,101,114,61,79,66,67,49,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115, 115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101, - 48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,51,58,48,48,48,48,45,102, - 102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110, - 116,105,102,105,101,114,61,83,68,68,49,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51, - 102,44,56,48,45,98,102,58,52,56,48,48,45,52,56,48,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32, - 109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102, - 102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102, - 102,102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, - 116,61,80,114,111,103,114,97,109,10,10,98,111,97,114,100,58,32,83,80,67,55,49,49,48,45,82,65,77,10,32,32, - 112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,83,80,67,55,49,49,48,10,32,32,32, - 32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,48,48,45,52,56, - 51,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,53,48,44,53,56,58,48,48,48,48,45,102,102, - 102,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45, - 51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,48,48, - 10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102, - 102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121, - 112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101,109, - 111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,10,32,32,32,32,109,101, - 109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32, - 32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102, - 102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,10,98,111,97,114,100,58,32,83,80,67,55,49,49,48,45,82, - 65,77,45,69,80,83,79,78,82,84,67,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105, - 101,114,61,83,80,67,55,49,49,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102, - 44,56,48,45,98,102,58,52,56,48,48,45,52,56,51,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,53,48,44,53,56,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102, - 32,109,97,115,107,61,48,120,56,48,48,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115, - 61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32, - 32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111, - 103,114,97,109,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101, - 110,116,61,68,97,116,97,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116, - 101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102, - 44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,114, - 116,99,32,109,97,110,117,102,97,99,116,117,114,101,114,61,69,112,115,111,110,10,32,32,32,32,109,97,112,32,97,100, - 100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,52,48,45,52,56,52,50,10,32,32,32,32, - 109,101,109,111,114,121,32,116,121,112,101,61,82,84,67,32,99,111,110,116,101,110,116,61,84,105,109,101,32,109,97,110, - 117,102,97,99,116,117,114,101,114,61,69,112,115,111,110,10,10,98,111,97,114,100,58,32,83,84,45,76,79,82,79,77, - 10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114, - 97,109,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58,56, - 48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,115,108,111,116,32,116,121,112,101, - 61,83,117,102,97,109,105,84,117,114,98,111,10,32,32,32,32,114,111,109,10,32,32,32,32,32,32,109,97,112,32,97, - 100,100,114,101,115,115,61,50,48,45,51,102,44,97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115, - 107,61,48,120,56,48,48,48,10,32,32,32,32,114,97,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101, - 115,115,61,54,48,45,54,102,44,101,48,45,101,102,58,48,48,48,48,45,102,102,102,102,10,32,32,115,108,111,116,32, - 116,121,112,101,61,83,117,102,97,109,105,84,117,114,98,111,10,32,32,32,32,114,111,109,10,32,32,32,32,32,32,109, - 97,112,32,97,100,100,114,101,115,115,61,52,48,45,53,102,44,99,48,45,100,102,58,48,48,48,48,45,102,102,102,102, - 32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32,114,97,109,10,32,32,32,32,32,32,109,97,112,32,97, - 100,100,114,101,115,115,61,55,48,45,55,100,44,102,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,10, + 48,48,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,49,44,102,48,45,102,49,58, + 54,48,48,48,45,55,102,102,102,44,101,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10, + 32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101, + 10,10,98,111,97,114,100,58,32,83,65,49,45,82,65,77,10,32,32,112,114,111,99,101,115,115,111,114,32,97,114,99, + 104,105,116,101,99,116,117,114,101,61,87,54,53,67,56,49,54,83,10,32,32,32,32,109,97,112,32,97,100,100,114,101, + 115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,50,50,48,48,45,50,51,102,102,10,32,32,32,32,109,99,117, + 10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56, + 48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,52,48,56,48,48,48,10,32,32,32,32,32,32,109,97,112, + 32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109, + 101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32, + 32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10, + 32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48, + 48,48,45,55,102,102,102,32,115,105,122,101,61,48,120,50,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,52,48,45,52,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,101,109,111,114,121, + 32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,73,110,116,101,114,110,97,108,10,32,32,32,32,32, + 32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,51,48,48,48,45,51,55, + 102,102,32,115,105,122,101,61,48,120,56,48,48,10,10,98,111,97,114,100,58,32,83,68,68,49,10,32,32,112,114,111, + 99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,83,68,68,49,10,32,32,32,32,109,97,112,32,97, + 100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,48,48,45,52,56,48,102,10,32,32,32, + 32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45, + 98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99, + 48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101, + 61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,10,98,111,97,114,100,58,32,83,68,68, + 49,45,82,65,77,10,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61, + 83,97,118,101,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102, + 58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,32,32,109,97,112,32,97, + 100,100,114,101,115,115,61,55,48,45,55,51,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48, + 48,48,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,83,68,68,49,10,32, + 32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,48,48,45, + 52,56,48,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48, + 48,45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,97,112,32,97, + 100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,32,32,109,101,109, + 111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,10,98,111, + 97,114,100,58,32,83,80,67,55,49,49,48,45,82,65,77,10,32,32,112,114,111,99,101,115,115,111,114,32,105,100,101, + 110,116,105,102,105,101,114,61,83,80,67,55,49,49,48,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61, + 48,48,45,51,102,44,56,48,45,98,102,58,52,56,48,48,45,52,56,51,102,10,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,53,48,44,53,56,58,48,48,48,48,45,102,102,102,102,10,32,32,32,32,109,99,117,10,32,32, + 32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,56,48,48,48, + 45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,48,48,10,32,32,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,99,48,48, + 48,48,48,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110, + 116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,79,77,32, + 99,111,110,116,101,110,116,61,68,97,116,97,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82,65,77, + 32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61, + 48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55,102,102,102,32,109,97,115,107,61,48,120,101,48,48, + 48,10,10,98,111,97,114,100,58,32,83,80,67,55,49,49,48,45,82,65,77,45,69,80,83,79,78,82,84,67,10,32, + 32,112,114,111,99,101,115,115,111,114,32,105,100,101,110,116,105,102,105,101,114,61,83,80,67,55,49,49,48,10,32,32, + 32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,52,56,48,48,45,52, + 56,51,102,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,53,48,44,53,56,58,48,48,48,48,45,102, + 102,102,102,10,32,32,32,32,109,99,117,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48, + 45,51,102,44,56,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,48, + 48,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,99,48,45,102,102,58,48,48,48,48,45,102, + 102,102,102,32,109,97,115,107,61,48,120,99,48,48,48,48,48,10,32,32,32,32,32,32,109,101,109,111,114,121,32,116, + 121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,32,32,109,101, + 109,111,114,121,32,116,121,112,101,61,82,79,77,32,99,111,110,116,101,110,116,61,68,97,116,97,10,32,32,32,32,109, + 101,109,111,114,121,32,116,121,112,101,61,82,65,77,32,99,111,110,116,101,110,116,61,83,97,118,101,10,32,32,32,32, + 32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56,48,45,98,102,58,54,48,48,48,45,55, + 102,102,102,32,109,97,115,107,61,48,120,101,48,48,48,10,32,32,114,116,99,32,109,97,110,117,102,97,99,116,117,114, + 101,114,61,69,112,115,111,110,10,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,48,48,45,51,102,44,56, + 48,45,98,102,58,52,56,52,48,45,52,56,52,50,10,32,32,32,32,109,101,109,111,114,121,32,116,121,112,101,61,82, + 84,67,32,99,111,110,116,101,110,116,61,84,105,109,101,32,109,97,110,117,102,97,99,116,117,114,101,114,61,69,112,115, + 111,110,10,10,98,111,97,114,100,58,32,83,84,45,76,79,82,79,77,10,32,32,109,101,109,111,114,121,32,116,121,112, + 101,61,82,79,77,32,99,111,110,116,101,110,116,61,80,114,111,103,114,97,109,10,32,32,32,32,109,97,112,32,97,100, + 100,114,101,115,115,61,48,48,45,49,102,44,56,48,45,57,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107, + 61,48,120,56,48,48,48,10,32,32,115,108,111,116,32,116,121,112,101,61,83,117,102,97,109,105,84,117,114,98,111,10, + 32,32,32,32,114,111,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,50,48,45,51,102,44, + 97,48,45,98,102,58,56,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10,32,32,32,32, + 114,97,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,54,48,45,54,102,44,101,48,45,101, + 102,58,48,48,48,48,45,102,102,102,102,10,32,32,115,108,111,116,32,116,121,112,101,61,83,117,102,97,109,105,84,117, + 114,98,111,10,32,32,32,32,114,111,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,52,48, + 45,53,102,44,99,48,45,100,102,58,48,48,48,48,45,102,102,102,102,32,109,97,115,107,61,48,120,56,48,48,48,10, + 32,32,32,32,114,97,109,10,32,32,32,32,32,32,109,97,112,32,97,100,100,114,101,115,115,61,55,48,45,55,100,44, + 102,48,45,102,102,58,48,48,48,48,45,102,102,102,102,10,10, }; const unsigned char iplrom[64] = { diff --git a/shaders/CRT-Royale.shader/bloom-approx.fs b/shaders/CRT-Royale.shader/bloom-approx.fs new file mode 100644 index 000000000..a56c09d6c --- /dev/null +++ b/shaders/CRT-Royale.shader/bloom-approx.fs @@ -0,0 +1,13973 @@ +#version 150 + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +in Vertex { + vec2 vTexCoord; + vec2 tex_uv; + vec2 blur_dxdy; + vec2 uv_scanline_step; + float estimated_viewport_size_x; + vec2 texture_size_inv; + vec2 tex_uv_to_pixel_scale; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +#define ORIG_LINEARIZEDvideo_size sourceSize[1].xy +#define ORIG_LINEARIZEDtexture_size sourceSize[1].xy +#define ORIG_LINEARIZED source[1] + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +/////////////////////////////// END VERTEX INCLUDES ///////////////////////////// + +////////////////////////////// FRAGMENT INCLUDES ////////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// output_size < video_size. +// 4.) output_size == video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (video_size/output_size)/texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(video_size/output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, output_size.x), [0, output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, output_size.x), [0, output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(video_size/output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +//#include "bloom-functions.h" + +//////////////////////////// BEGIN BLOOM-FUNCTIONS /////////////////////////// + +#ifndef BLOOM_FUNCTIONS_H +#define BLOOM_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These utility functions and constants help several passes determine the +// size and center texel weight of the phosphor bloom in a uniform manner. + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// We need to calculate the correct blur sigma using some .cgp constants: +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// output_size < video_size. +// 4.) output_size == video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (video_size/output_size)/texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(video_size/output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, output_size.x), [0, output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, output_size.x), [0, output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(video_size/output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +/////////////////////////////// BLOOM CONSTANTS ////////////////////////////// + +// Compute constants with manual inlines of the functions below: +static const float bloom_diff_thresh = 1.0/256.0; + + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +inline float get_absolute_scale_blur_sigma(const float thresh) +{ + // Requires: 1.) min_expected_triads must be a global float. The number + // of horizontal phosphor triads in the final image must be + // >= min_allowed_viewport_triads.x for realistic results. + // 2.) bloom_approx_scale_x must be a global float equal to the + // absolute horizontal scale of BLOOM_APPROX. + // 3.) bloom_approx_scale_x/min_allowed_viewport_triads.x + // should be <= 1.1658025090 to keep the final result < + // 0.62666015625 (the largest sigma ensuring the largest + // unused texel weight stays < 1.0/256.0 for a 3x3 blur). + // 4.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum Gaussian sigma that will blur the pass + // output as much as it would have taken to blur away + // bloom_approx_scale_x horizontal phosphor triads. + // Description: + // BLOOM_APPROX should look like a downscaled phosphor blur. Ideally, we'd + // use the same blur sigma as the actual phosphor bloom and scale it down + // to the current resolution with (bloom_approx_scale_x/viewport_size_x), but + // we don't know the viewport size in this pass. Instead, we'll blur as + // much as it would take to blur away min_allowed_viewport_triads.x. This + // will blur "more than necessary" if the user actually uses more triads, + // but that's not terrible either, because blurring a constant fraction of + // the viewport may better resemble a true optical bloom anyway (since the + // viewport will generally be about the same fraction of each player's + // field of view, regardless of screen size and resolution). + // Assume an extremely large viewport size for asymptotic results. + return bloom_approx_scale_x/max_viewport_size_x * + get_min_sigma_to_blur_triad( + max_viewport_size_x/min_allowed_viewport_triads.x, thresh); +} + +inline float get_center_weight(const float sigma) +{ + // Given a Gaussian blur sigma, get the blur weight for the center texel. + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return get_fast_gaussian_weight_sum_inv(sigma); + #else + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + // Note: If the implementation uses a smaller blur than the max allowed, + // the worst case scenario is that the center weight will be overestimated, + // so we'll put a bit more energy into the brightpass...no huge deal. + // Then again, if the implementation uses a larger blur than the max + // "allowed" because of dynamic branching, the center weight could be + // underestimated, which is more of a problem...consider always using + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // 43x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + + w11 + w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + // 31x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + + w8 + w9 + w10 + w11 + w12 + w13 + w14 + w15)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + // 25x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + // 17x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + #else + // 9x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + const float center_weight = weight_sum_inv * weight_sum_inv; + return center_weight; + #endif +} + +inline float3 tex2DblurNfast(const sampler2D texture, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // If sigma is static, we can safely branch and use the smallest blur + // that's big enough. Ignore #define hints, because we'll only use a + // large blur if we actually need it, and the branches cost nothing. + #ifndef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #else + // It's still worth branching if the profile supports dynamic branches: + // It's much faster than using a hugely excessive blur, but each branch + // eats ~1% FPS. + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #endif + #endif + // Failed optimization notes: + // I originally created a same-size mipmapped 5-tap separable blur10 that + // could handle any sigma by reaching into lower mip levels. It was + // as fast as blur25fast for runtime sigmas and a tad faster than + // blur31fast for static sigmas, but mipmapping two viewport-size passes + // ate 10% of FPS across all codepaths, so it wasn't worth it. + #ifdef PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + if(sigma <= blur9_std_dev) + { + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur17_std_dev) + { + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur25_std_dev) + { + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur31_std_dev) + { + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + } + else + { + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + } + #else + // If we can't afford to branch, we can only guess at what blur + // size we need. Therefore, use the largest blur allowed. + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + #else + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + #endif // PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE +} + +inline float get_bloom_approx_sigma(const float output_size_x_runtime, + const float estimated_viewport_size_x) +{ + // Requires: 1.) output_size_x_runtime == BLOOM_APPROX.output_size.x. + // This is included for dynamic codepaths just in case the + // following two globals are incorrect: + // 2.) bloom_approx_size_x_for_skip should == the same + // if PHOSPHOR_BLOOM_FAKE is #defined + // 3.) bloom_approx_size_x should == the same otherwise + // Returns: For gaussian4x4, return a dynamic small bloom sigma that's + // as close to optimal as possible given available information. + // For blur3x3, return the a static small bloom sigma that + // works well for typical cases. Otherwise, we're using simple + // bilinear filtering, so use static calculations. + // Assume the default static value. This is a compromise that ensures + // typical triads are blurred, even if unusually large ones aren't. + static const float mask_num_triads_static = + max(min_allowed_viewport_triads.x, mask_num_triads_desired_static); + const float mask_num_triads_from_size = + estimated_viewport_size_x/mask_triad_size_desired; + const float mask_num_triads_runtime = max(min_allowed_viewport_triads.x, + lerp(mask_num_triads_from_size, mask_num_triads_desired, + mask_specify_num_triads)); + // Assume an extremely large viewport size for asymptotic results: + static const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + if(bloom_approx_filter > 1.5) // 4x4 true Gaussian resize + { + // Use the runtime num triads and output size: + const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_runtime; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_runtime/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // account for the Gaussian scanline sigma from the last pass too. + // The bloom will be too wide horizontally but tall enough vertically. + return length(float2(bloom_approx_sigma, beam_max_sigma)); + } + else // 3x3 blur resize (the bilinear resize doesn't need a sigma) + { + // We're either using blur3x3 or bilinear filtering. The biggest + // reason to choose blur3x3 is to avoid dynamic weights, so use a + // static calculation. + #ifdef PHOSPHOR_BLOOM_FAKE + static const float output_size_x_static = + bloom_approx_size_x_for_fake; + #else + static const float output_size_x_static = bloom_approx_size_x; + #endif + static const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_static; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_static/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // try accounting for the Gaussian scanline sigma from the last pass + // too; use the static default value: + return length(float2(bloom_approx_sigma, beam_max_sigma_static)); + } +} + +inline float get_final_bloom_sigma(const float bloom_sigma_runtime) +{ + // Requires: 1.) bloom_sigma_runtime is a precalculated sigma that's + // optimal for the [known] triad size. + // 2.) Call this from a fragment shader (not a vertex shader), + // or blurring with static sigmas won't be constant-folded. + // Returns: Return the optimistic static sigma if the triad size is + // known at compile time. Otherwise return the optimal runtime + // sigma (10% slower) or an implementation-specific compromise + // between an optimistic or pessimistic static sigma. + // Notes: Call this from the fragment shader, NOT the vertex shader, + // so static sigmas can be constant-folded! + const float bloom_sigma_optimistic = get_min_sigma_to_blur_triad( + mask_triad_size_desired_static, bloom_diff_thresh); + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return bloom_sigma_runtime; + #else + // Overblurring looks as bad as underblurring, so assume average-size + // triads, not worst-case huge triads: + return bloom_sigma_optimistic; + #endif +} + + +#endif // BLOOM_FUNCTIONS_H + +//////////////////////////// END BLOOM-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +/////////////////////////////////// HELPERS ////////////////////////////////// + +float3 tex2Dresize_gaussian4x4(sampler2D tex, float2 tex_uv, float2 dxdy, float2 tex_size, float2 texture_size_inv, float2 tex_uv_to_pixel_scale, float sigma) +{ + // Requires: 1.) All requirements of gamma-management.h must be satisfied! + // 2.) filter_linearN must == "true" in your .cgp preset. + // 3.) mipmap_inputN must == "true" in your .cgp preset if + // output_size << SRC.video_size. + // 4.) dxdy should contain the uv pixel spacing: + // dxdy = max(float2(1.0), + // SRC.video_size/output_size)/SRC.texture_size; + // 5.) texture_size == SRC.texture_size + // 6.) texture_size_inv == float2(1.0)/SRC.texture_size + // 7.) tex_uv_to_pixel_scale == output_size * + // SRC.texture_size / SRC.video_size; + // 8.) sigma is the desired Gaussian standard deviation, in + // terms of output pixels. It should be < ~0.66171875 to + // ensure the first unused sample (outside the 4x4 box) has + // a weight < 1.0/256.0. + // Returns: A true 4x4 Gaussian resize of the input. + // Description: + // Given correct inputs, this Gaussian resizer samples 4 pixel locations + // along each downsized dimension and/or 4 texel locations along each + // upsized dimension. It computes dynamic weights based on the pixel-space + // distance of each sample from the destination pixel. It is arbitrarily + // resizable and higher quality than tex2Dblur3x3_resize, but it's slower. + // TODO: Move this to a more suitable file once there are others like it. + const float denom_inv = 0.5/(sigma*sigma); + // We're taking 4x4 samples, and we're snapping to texels for upsizing. + // Find texture coords for sample 5 (second row, second column): + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_uv = prev_texel * texture_size_inv; + const float2 snap = float2((dxdy.x <= texture_size_inv.x), (dxdy.y <= texture_size_inv.y)); + const float2 sample5_downsize_uv = tex_uv - 0.5 * dxdy; + const float2 sample5_uv = lerp(sample5_downsize_uv, prev_texel_uv, snap); + // Compute texture coords for other samples: + const float2 dx = float2(dxdy.x, 0.0); + const float2 sample0_uv = sample5_uv - dxdy; + const float2 sample10_uv = sample5_uv + dxdy; + const float2 sample15_uv = sample5_uv + 2.0 * dxdy; + const float2 sample1_uv = sample0_uv + dx; + const float2 sample2_uv = sample0_uv + 2.0 * dx; + const float2 sample3_uv = sample0_uv + 3.0 * dx; + const float2 sample4_uv = sample5_uv - dx; + const float2 sample6_uv = sample5_uv + dx; + const float2 sample7_uv = sample5_uv + 2.0 * dx; + const float2 sample8_uv = sample10_uv - 2.0 * dx; + const float2 sample9_uv = sample10_uv - dx; + const float2 sample11_uv = sample10_uv + dx; + const float2 sample12_uv = sample15_uv - 3.0 * dx; + const float2 sample13_uv = sample15_uv - 2.0 * dx; + const float2 sample14_uv = sample15_uv - dx; + // Load each sample: + float3 sample0 = tex2D_linearize(tex, sample0_uv).rgb; + float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + float3 sample2 = tex2D_linearize(tex, dx).rgb; + float3 sample3 = tex2D_linearize(tex, sample3_uv).rgb; + float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + float3 sample5 = tex2D_linearize(tex, sample5_uv).rgb; + float3 sample6 = tex2D_linearize(tex, sample6_uv).rgb; + float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + float3 sample8 = tex2D_linearize(tex, sample8_uv).rgb; + float3 sample9 = tex2D_linearize(tex, sample9_uv).rgb; + float3 sample10 = tex2D_linearize(tex, sample10_uv).rgb; + float3 sample11 = tex2D_linearize(tex, sample11_uv).rgb; + float3 sample12 = tex2D_linearize(tex, sample12_uv).rgb; + float3 sample13 = tex2D_linearize(tex, sample13_uv).rgb; + float3 sample14 = tex2D_linearize(tex, sample14_uv).rgb; + float3 sample15 = tex2D_linearize(tex, sample15_uv).rgb; + // Compute destination pixel offsets for each sample: + const float2 dest_pixel = tex_uv * tex_uv_to_pixel_scale; + const float2 sample0_offset = sample0_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample1_offset = sample1_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample2_offset = sample2_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample3_offset = sample3_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample4_offset = sample4_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample5_offset = sample5_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample6_offset = sample6_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample7_offset = sample7_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample8_offset = sample8_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample9_offset = sample9_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample10_offset = sample10_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample11_offset = sample11_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample12_offset = sample12_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample13_offset = sample13_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample14_offset = sample14_uv * tex_uv_to_pixel_scale - dest_pixel; + const float2 sample15_offset = sample15_uv * tex_uv_to_pixel_scale - dest_pixel; + // Compute Gaussian sample weights: + const float w0 = exp(-LENGTH_SQ(sample0_offset) * denom_inv); + const float w1 = exp(-LENGTH_SQ(sample1_offset) * denom_inv); + const float w2 = exp(-LENGTH_SQ(sample2_offset) * denom_inv); + const float w3 = exp(-LENGTH_SQ(sample3_offset) * denom_inv); + const float w4 = exp(-LENGTH_SQ(sample4_offset) * denom_inv); + const float w5 = exp(-LENGTH_SQ(sample5_offset) * denom_inv); + const float w6 = exp(-LENGTH_SQ(sample6_offset) * denom_inv); + const float w7 = exp(-LENGTH_SQ(sample7_offset) * denom_inv); + const float w8 = exp(-LENGTH_SQ(sample8_offset) * denom_inv); + const float w9 = exp(-LENGTH_SQ(sample9_offset) * denom_inv); + const float w10 = exp(-LENGTH_SQ(sample10_offset) * denom_inv); + const float w11 = exp(-LENGTH_SQ(sample11_offset) * denom_inv); + const float w12 = exp(-LENGTH_SQ(sample12_offset) * denom_inv); + const float w13 = exp(-LENGTH_SQ(sample13_offset) * denom_inv); + const float w14 = exp(-LENGTH_SQ(sample14_offset) * denom_inv); + const float w15 = exp(-LENGTH_SQ(sample15_offset) * denom_inv); + const float weight_sum_inv = 1.0/( + w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + + w8 +w9 + w10 + w11 + w12 + w13 + w14 + w15); + // Weight and sum the samples: + const float3 sum = w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15; + return sum * weight_sum_inv; +} + +void main() { + // Would a viewport-relative size work better for this pass? (No.) + // PROS: + // 1.) Instead of writing an absolute size to user-cgp-constants.h, we'd + // write a viewport scale. That number could be used to directly scale + // the viewport-resolution bloom sigma and/or triad size to a smaller + // scale. This way, we could calculate an optimal dynamic sigma no + // matter how the dot pitch is specified. + // CONS: + // 1.) Texel smearing would be much worse at small viewport sizes, but + // performance would be much worse at large viewport sizes, so there + // would be no easy way to calculate a decent scale. + // 2.) Worse, we could no longer get away with using a constant-size blur! + // Instead, we'd have to face all the same difficulties as the real + // phosphor bloom, which requires static #ifdefs to decide the blur + // size based on the expected triad size...a dynamic value. + // 3.) Like the phosphor bloom, we'd have less control over making the blur + // size correct for an optical blur. That said, we likely overblur (to + // maintain brightness) more than the eye would do by itself: 20/20 + // human vision distinguishes ~1 arc minute, or 1/60 of a degree. The + // highest viewing angle recommendation I know of is THX's 40.04 degree + // recommendation, at which 20/20 vision can distinguish about 2402.4 + // lines. Assuming the "TV lines" definition, that means 1201.2 + // distinct light lines and 1201.2 distinct dark lines can be told + // apart, i.e. 1201.2 pairs of lines. This would correspond to 1201.2 + // pairs of alternating lit/unlit phosphors, so 2402.4 phosphors total + // (if they're alternately lit). That's a max of 800.8 triads. Using + // a more popular 30 degree viewing angle recommendation, 20/20 vision + // can distinguish 1800 lines, or 600 triads of alternately lit + // phosphors. In contrast, we currently blur phosphors all the way + // down to 341.3 triads to ensure full brightness. + // 4.) Realistically speaking, we're usually just going to use bilinear + // filtering in this pass anyway, but it only works well to limit + // bandwidth if it's done at a small constant scale. + + // Get the constants we need to sample: +// const sampler2D texture = ORIG_LINEARIZED.texture; +// const float2 tex_uv = tex_uv; +// const float2 blur_dxdy = blur_dxdy; + const float2 texture_size_ = ORIG_LINEARIZEDtexture_size; +// const float2 texture_size_inv = texture_size_inv; +// const float2 tex_uv_to_pixel_scale = tex_uv_to_pixel_scale; + float2 tex_uv_r, tex_uv_g, tex_uv_b; + + if(beam_misconvergence) + { + const float2 uv_scanline_step = uv_scanline_step; + const float2 convergence_offsets_r = get_convergence_offsets_r_vector(); + const float2 convergence_offsets_g = get_convergence_offsets_g_vector(); + const float2 convergence_offsets_b = get_convergence_offsets_b_vector(); + tex_uv_r = tex_uv - convergence_offsets_r * uv_scanline_step; + tex_uv_g = tex_uv - convergence_offsets_g * uv_scanline_step; + tex_uv_b = tex_uv - convergence_offsets_b * uv_scanline_step; + } + // Get the blur sigma: + const float bloom_approx_sigma = get_bloom_approx_sigma(output_size.x, + estimated_viewport_size_x); + + // Sample the resized and blurred texture, and apply convergence offsets if + // necessary. Applying convergence offsets here triples our samples from + // 16/9/1 to 48/27/3, but faster and easier than sampling BLOOM_APPROX and + // HALATION_BLUR 3 times at full resolution every time they're used. + float3 color_r, color_g, color_b, color; + if(bloom_approx_filter > 1.5) + { + // Use a 4x4 Gaussian resize. This is slower but technically correct. + if(beam_misconvergence) + { + color_r = tex2Dresize_gaussian4x4(ORIG_LINEARIZED, tex_uv_r, + blur_dxdy, texture_size_, texture_size_inv, + tex_uv_to_pixel_scale, bloom_approx_sigma); + color_g = tex2Dresize_gaussian4x4(ORIG_LINEARIZED, tex_uv_g, + blur_dxdy, texture_size_, texture_size_inv, + tex_uv_to_pixel_scale, bloom_approx_sigma); + color_b = tex2Dresize_gaussian4x4(ORIG_LINEARIZED, tex_uv_b, + blur_dxdy, texture_size_, texture_size_inv, + tex_uv_to_pixel_scale, bloom_approx_sigma); + } + else + { + color = tex2Dresize_gaussian4x4(ORIG_LINEARIZED, tex_uv, + blur_dxdy, texture_size_, texture_size_inv, + tex_uv_to_pixel_scale, bloom_approx_sigma); + } + } + else if(bloom_approx_filter > 0.5) + { + // Use a 3x3 resize blur. This is the softest option, because we're + // blurring already blurry bilinear samples. It doesn't play quite as + // nicely with convergence offsets, but it has its charms. + if(beam_misconvergence) + { + color_r = tex2Dblur3x3resize(ORIG_LINEARIZED, tex_uv_r, + blur_dxdy, bloom_approx_sigma); + color_g = tex2Dblur3x3resize(ORIG_LINEARIZED, tex_uv_g, + blur_dxdy, bloom_approx_sigma); + color_b = tex2Dblur3x3resize(ORIG_LINEARIZED, tex_uv_b, + blur_dxdy, bloom_approx_sigma); + } + else + { + color = tex2Dblur3x3resize(ORIG_LINEARIZED, tex_uv, blur_dxdy); + } + } + else + { + // Use bilinear sampling. This approximates a 4x4 Gaussian resize MUCH + // better than tex2Dblur3x3_resize for the very small sigmas we're + // likely to use at small output resolutions. (This estimate becomes + // too sharp above ~400x300, but the blurs break down above that + // resolution too, unless min_allowed_viewport_triads is high enough to + // keep bloom_approx_scale_x/min_allowed_viewport_triads < ~1.1658025.) + if(beam_misconvergence) + { + color_r = tex2D_linearize(ORIG_LINEARIZED, tex_uv_r).rgb; + color_g = tex2D_linearize(ORIG_LINEARIZED, tex_uv_g).rgb; + color_b = tex2D_linearize(ORIG_LINEARIZED, tex_uv_b).rgb; + } + else + { + color = tex2D_linearize(ORIG_LINEARIZED, tex_uv).rgb; + } + } + // Pack the colors from the red/green/blue beams into a single vector: + if(beam_misconvergence) + { + color = float3(color_r.r, color_g.g, color_b.b); + } + // Encode and output the blurred image: + FragColor = encode_output(float4(tex2D_linearize(ORIG_LINEARIZED, tex_uv))); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/bloom-approx.vs b/shaders/CRT-Royale.shader/bloom-approx.vs new file mode 100644 index 000000000..e4faac1e6 --- /dev/null +++ b/shaders/CRT-Royale.shader/bloom-approx.vs @@ -0,0 +1,5859 @@ +#version 150 + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 tex_uv; + vec2 blur_dxdy; + vec2 uv_scanline_step; + float estimated_viewport_size_x; + vec2 texture_size_inv; + vec2 tex_uv_to_pixel_scale; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; +uniform int phase; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +#define ORIG_LINEARIZEDvideo_size sourceSize[1].xy +#define ORIG_LINEARIZEDtexture_size sourceSize[1].xy + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +/////////////////////////////// END VERTEX INCLUDES ///////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + +void main() { + gl_Position = position; + vTexCoord = texCoord; + const float2 video_uv = vTexCoord * texture_size/video_size; + tex_uv = video_uv * ORIG_LINEARIZEDvideo_size / + ORIG_LINEARIZEDtexture_size; + // The last pass (vertical scanlines) had a viewport y scale, so we can + // use it to calculate a better runtime sigma: + estimated_viewport_size_x = + video_size.y * geom_aspect_ratio_x/geom_aspect_ratio_y; + + // Get the uv sample distance between output pixels. We're using a resize + // blur, so arbitrary upsizing will be acceptable if filter_linearN = + // "true," and arbitrary downsizing will be acceptable if mipmap_inputN = + // "true" too. The blur will be much more accurate if a true 4x4 Gaussian + // resize is used instead of tex2Dblur3x3_resize (which samples between + // texels even for upsizing). + const float2 dxdy_min_scale = ORIG_LINEARIZEDvideo_size/output_size; + const float2 texture_size_inv = float2(1.0, 1.0)/ORIG_LINEARIZEDtexture_size; + if(bloom_approx_filter > 1.5) // 4x4 true Gaussian resize + { + // For upsizing, we'll snap to texels and sample the nearest 4. + const float2 dxdy_scale = max(dxdy_min_scale, float2(1.0, 1.0)); + blur_dxdy = dxdy_scale * texture_size_inv; + } + else + { + const float2 dxdy_scale = dxdy_min_scale; + blur_dxdy = dxdy_scale * texture_size_inv; + } + // tex2Dresize_gaussian4x4 needs to know a bit more than the other filters: + tex_uv_to_pixel_scale = output_size * + ORIG_LINEARIZEDtexture_size / ORIG_LINEARIZEDvideo_size; + //texture_size_inv = texture_size_inv; + + // Detecting interlacing again here lets us apply convergence offsets in + // this pass. il_step_multiple contains the (texel, scanline) step + // multiple: 1 for progressive, 2 for interlaced. + const float2 orig_video_size = ORIG_LINEARIZEDvideo_size; + const float y_step = 1.0 + float(is_interlaced(orig_video_size.y)); + const float2 il_step_multiple = float2(1.0, y_step); + // Get the uv distance between (texels, same-field scanlines): + uv_scanline_step = il_step_multiple / ORIG_LINEARIZEDtexture_size; +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/bloom-horizontal-reconstitute.fs b/shaders/CRT-Royale.shader/bloom-horizontal-reconstitute.fs new file mode 100644 index 000000000..7750152cf --- /dev/null +++ b/shaders/CRT-Royale.shader/bloom-horizontal-reconstitute.fs @@ -0,0 +1,7240 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +// Dunno why this stuff didn't want to function in the vertex, but whatever +in Vertex { + vec2 vTexCoord; +// vec2 video_uv; +// vec2 scanline_tex_uv; +// vec2 halation_tex_uv; +// vec2 brightpass_tex_uv; +// vec2 bloom_tex_uv; + vec2 bloom_dxdy; + float bloom_sigma_runtime; +}; + + vec2 video_uv = vTexCoord; + vec2 scanline_tex_uv = vTexCoord; + vec2 halation_tex_uv = vTexCoord; + vec2 brightpass_tex_uv = vTexCoord; + vec2 bloom_tex_uv = vTexCoord; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define bloom_texture source[0] + +#define MASKED_SCANLINEStexture source[2] +#define MASKED_SCANLINEStexture_size sourceSize[2].xy +#define MASKED_SCANLINESvideo_size sourceSize[2].xy +#define HALATION_BLURtexture source[5] +#define HALATION_BLURtexture_size sourceSize[5].xy +#define HALATION_BLURvideo_size sourceSize[5].xy +#define BRIGHTPASStexture source[1] +#define BRIGHTPASStexture_size sourceSize[1].xy +#define BRIGHTPASSvideo_size sourceSize[1].xy + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-params.h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); +const float bloom_diff_thresh_ = 1.0/256.0; + +/////////////////////////// BEGIN FRAGMENT-INCLUDES /////////////////////////// + +//#include "bloom-functions.h" + +//////////////////////////// BEGIN BLOOM-FUNCTIONS /////////////////////////// + +#ifndef BLOOM_FUNCTIONS_H +#define BLOOM_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These utility functions and constants help several passes determine the +// size and center texel weight of the phosphor bloom in a uniform manner. + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// We need to calculate the correct blur sigma using some .cgp constants: +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// output_size < video_size. +// 4.) output_size == video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (video_size/output_size)/texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(video_size/output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, output_size.x), [0, output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, output_size.x), [0, output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(video_size/output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +/////////////////////////////// BLOOM CONSTANTS ////////////////////////////// + +// Compute constants with manual inlines of the functions below: +static const float bloom_diff_thresh = 1.0/256.0; + + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +inline float get_absolute_scale_blur_sigma(const float thresh) +{ + // Requires: 1.) min_expected_triads must be a global float. The number + // of horizontal phosphor triads in the final image must be + // >= min_allowed_viewport_triads.x for realistic results. + // 2.) bloom_approx_scale_x must be a global float equal to the + // absolute horizontal scale of BLOOM_APPROX. + // 3.) bloom_approx_scale_x/min_allowed_viewport_triads.x + // should be <= 1.1658025090 to keep the final result < + // 0.62666015625 (the largest sigma ensuring the largest + // unused texel weight stays < 1.0/256.0 for a 3x3 blur). + // 4.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum Gaussian sigma that will blur the pass + // output as much as it would have taken to blur away + // bloom_approx_scale_x horizontal phosphor triads. + // Description: + // BLOOM_APPROX should look like a downscaled phosphor blur. Ideally, we'd + // use the same blur sigma as the actual phosphor bloom and scale it down + // to the current resolution with (bloom_approx_scale_x/viewport_size_x), but + // we don't know the viewport size in this pass. Instead, we'll blur as + // much as it would take to blur away min_allowed_viewport_triads.x. This + // will blur "more than necessary" if the user actually uses more triads, + // but that's not terrible either, because blurring a constant fraction of + // the viewport may better resemble a true optical bloom anyway (since the + // viewport will generally be about the same fraction of each player's + // field of view, regardless of screen size and resolution). + // Assume an extremely large viewport size for asymptotic results. + return bloom_approx_scale_x/max_viewport_size_x * + get_min_sigma_to_blur_triad( + max_viewport_size_x/min_allowed_viewport_triads.x, thresh); +} + +inline float get_center_weight(const float sigma) +{ + // Given a Gaussian blur sigma, get the blur weight for the center texel. + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return get_fast_gaussian_weight_sum_inv(sigma); + #else + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + // Note: If the implementation uses a smaller blur than the max allowed, + // the worst case scenario is that the center weight will be overestimated, + // so we'll put a bit more energy into the brightpass...no huge deal. + // Then again, if the implementation uses a larger blur than the max + // "allowed" because of dynamic branching, the center weight could be + // underestimated, which is more of a problem...consider always using + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // 43x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + + w11 + w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + // 31x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + + w8 + w9 + w10 + w11 + w12 + w13 + w14 + w15)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + // 25x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + // 17x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + #else + // 9x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + const float center_weight = weight_sum_inv * weight_sum_inv; + return center_weight; + #endif +} + +inline float3 tex2DblurNfast(const sampler2D texture, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // If sigma is static, we can safely branch and use the smallest blur + // that's big enough. Ignore #define hints, because we'll only use a + // large blur if we actually need it, and the branches cost nothing. + #ifndef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #else + // It's still worth branching if the profile supports dynamic branches: + // It's much faster than using a hugely excessive blur, but each branch + // eats ~1% FPS. + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #endif + #endif + // Failed optimization notes: + // I originally created a same-size mipmapped 5-tap separable blur10 that + // could handle any sigma by reaching into lower mip levels. It was + // as fast as blur25fast for runtime sigmas and a tad faster than + // blur31fast for static sigmas, but mipmapping two viewport-size passes + // ate 10% of FPS across all codepaths, so it wasn't worth it. + #ifdef PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + if(sigma <= blur9_std_dev) + { + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur17_std_dev) + { + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur25_std_dev) + { + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur31_std_dev) + { + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + } + else + { + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + } + #else + // If we can't afford to branch, we can only guess at what blur + // size we need. Therefore, use the largest blur allowed. + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + #else + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + #endif // PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE +} + +inline float get_bloom_approx_sigma(const float output_size_x_runtime, + const float estimated_viewport_size_x) +{ + // Requires: 1.) output_size_x_runtime == BLOOM_APPROX.output_size.x. + // This is included for dynamic codepaths just in case the + // following two globals are incorrect: + // 2.) bloom_approx_size_x_for_skip should == the same + // if PHOSPHOR_BLOOM_FAKE is #defined + // 3.) bloom_approx_size_x should == the same otherwise + // Returns: For gaussian4x4, return a dynamic small bloom sigma that's + // as close to optimal as possible given available information. + // For blur3x3, return the a static small bloom sigma that + // works well for typical cases. Otherwise, we're using simple + // bilinear filtering, so use static calculations. + // Assume the default static value. This is a compromise that ensures + // typical triads are blurred, even if unusually large ones aren't. + static const float mask_num_triads_static = + max(min_allowed_viewport_triads.x, mask_num_triads_desired_static); + const float mask_num_triads_from_size = + estimated_viewport_size_x/mask_triad_size_desired; + const float mask_num_triads_runtime = max(min_allowed_viewport_triads.x, + lerp(mask_num_triads_from_size, mask_num_triads_desired, + mask_specify_num_triads)); + // Assume an extremely large viewport size for asymptotic results: + static const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + if(bloom_approx_filter > 1.5) // 4x4 true Gaussian resize + { + // Use the runtime num triads and output size: + const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_runtime; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_runtime/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // account for the Gaussian scanline sigma from the last pass too. + // The bloom will be too wide horizontally but tall enough vertically. + return length(float2(bloom_approx_sigma, beam_max_sigma)); + } + else // 3x3 blur resize (the bilinear resize doesn't need a sigma) + { + // We're either using blur3x3 or bilinear filtering. The biggest + // reason to choose blur3x3 is to avoid dynamic weights, so use a + // static calculation. + #ifdef PHOSPHOR_BLOOM_FAKE + static const float output_size_x_static = + bloom_approx_size_x_for_fake; + #else + static const float output_size_x_static = bloom_approx_size_x; + #endif + static const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_static; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_static/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // try accounting for the Gaussian scanline sigma from the last pass + // too; use the static default value: + return length(float2(bloom_approx_sigma, beam_max_sigma_static)); + } +} + +inline float get_final_bloom_sigma(const float bloom_sigma_runtime) +{ + // Requires: 1.) bloom_sigma_runtime is a precalculated sigma that's + // optimal for the [known] triad size. + // 2.) Call this from a fragment shader (not a vertex shader), + // or blurring with static sigmas won't be constant-folded. + // Returns: Return the optimistic static sigma if the triad size is + // known at compile time. Otherwise return the optimal runtime + // sigma (10% slower) or an implementation-specific compromise + // between an optimistic or pessimistic static sigma. + // Notes: Call this from the fragment shader, NOT the vertex shader, + // so static sigmas can be constant-folded! + const float bloom_sigma_optimistic = get_min_sigma_to_blur_triad( + mask_triad_size_desired_static, bloom_diff_thresh); + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return bloom_sigma_runtime; + #else + // Overblurring looks as bad as underblurring, so assume average-size + // triads, not worst-case huge triads: + return bloom_sigma_optimistic; + #endif +} + + +#endif // BLOOM_FUNCTIONS_H + +//////////////////////////// END BLOOM-FUNCTIONS /////////////////////////// + +/////////////////////////// END FRAGMENT-INCLUDES ////////////////////////// + +void main() { + // Blur the vertically blurred brightpass horizontally by 9/17/25/43x: + const float bloom_sigma = get_final_bloom_sigma(bloom_sigma_runtime); + const float3 blurred_brightpass = tex2DblurNfast(bloom_texture, + bloom_tex_uv, bloom_dxdy, bloom_sigma); + + // Sample the masked scanlines. Alpha contains the auto-dim factor: + const float3 intensity_dim = + tex2D_linearize(MASKED_SCANLINEStexture, scanline_tex_uv).rgb; + const float auto_dim_factor = levels_autodim_temp; + const float undim_factor = 1.0/auto_dim_factor; + + // Calculate the mask dimpass, add it to the blurred brightpass, and + // undim (from scanline auto-dim) and amplify (from mask dim) the result: + const float mask_amplify = get_mask_amplify(); + const float3 brightpass = tex2D_linearize(BRIGHTPASStexture, + brightpass_tex_uv).rgb; + const float3 dimpass = intensity_dim - brightpass; + const float3 phosphor_bloom = (dimpass + blurred_brightpass) * + mask_amplify * undim_factor * levels_contrast; + + // Sample the halation texture, and let some light bleed into refractive + // diffusion. Conceptually this occurs before the phosphor bloom, but + // adding it in earlier passes causes black crush in the diffusion colors. + const float3 diffusion_color = levels_contrast * tex2D_linearize( + HALATION_BLURtexture, halation_tex_uv).rgb; + const float3 final_bloom = lerp(phosphor_bloom, + diffusion_color, diffusion_weight); + + // Encode and output the bloomed image: + FragColor = encode_output(float4(final_bloom, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/bloom-horizontal-reconstitute.vs b/shaders/CRT-Royale.shader/bloom-horizontal-reconstitute.vs new file mode 100644 index 000000000..5d9ad005a --- /dev/null +++ b/shaders/CRT-Royale.shader/bloom-horizontal-reconstitute.vs @@ -0,0 +1,6570 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +in vec4 position; +in vec2 texCoord; + +// These things didn't want to function in the vertex, so I just commented them +out Vertex { + vec2 vTexCoord; +// vec2 video_uv; +// vec2 scanline_tex_uv; +// vec2 halation_tex_uv; +// vec2 brightpass_tex_uv; +// vec2 bloom_tex_uv; + vec2 bloom_dxdy; + float bloom_sigma_runtime; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define MASKED_SCANLINEStexture source[2] +#define MASKED_SCANLINEStexture_size sourceSize[2].xy +#define MASKED_SCANLINESvideo_size sourceSize[2].xy +#define HALATION_BLURtexture source[5] +#define HALATION_BLURtexture_size sourceSize[5].xy +#define HALATION_BLURvideo_size sourceSize[5].xy +#define BRIGHTPASStexture source[1] +#define BRIGHTPASStexture_size sourceSize[1].xy +#define BRIGHTPASSvideo_size sourceSize[1].xy + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-params.h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +/////////////////////////////// VERTEX INCLUDES ////////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +////////////////////////////// END VERTEX-INCLUDES ////////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); +const float bloom_diff_thresh_ = 1.0/256.0; + +// copied from bloom-functions.h +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +void main() { + gl_Position = position; + vTexCoord = texCoord * 1.0001; + float2 tex_uv = vTexCoord.xy; + +// These things keep causing weird behavior and they're not needed except for NPOT, so... +/* // Our various input textures use different coords: + const float2 video_uv = tex_uv;// * texture_size/video_size; + video_uv = video_uv; + scanline_tex_uv = video_uv;// * MASKED_SCANLINESvideo_size / + MASKED_SCANLINEStexture_size; + halation_tex_uv = video_uv;// * HALATION_BLURvideo_size / + HALATION_BLURtexture_size; + brightpass_tex_uv = video_uv;// * BRIGHTPASSvideo_size / + BRIGHTPASStexture_size; + bloom_tex_uv = tex_uv; +*/ + // We're horizontally blurring the bloom input (vertically blurred + // brightpass). Get the uv distance between output pixels / input texels + // in the horizontal direction (this pass must NOT resize): + bloom_dxdy = float2(1.0/texture_size.x, 0.0); + + // Calculate a runtime bloom_sigma in case it's needed: + const float mask_tile_size_x = get_resized_mask_tile_size( + output_size, output_size * mask_resize_viewport_scale, false).x; + bloom_sigma_runtime = get_min_sigma_to_blur_triad( + mask_tile_size_x / mask_triads_per_tile, bloom_diff_thresh_); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/bloom-vertical.fs b/shaders/CRT-Royale.shader/bloom-vertical.fs new file mode 100644 index 000000000..4c37eee1e --- /dev/null +++ b/shaders/CRT-Royale.shader/bloom-vertical.fs @@ -0,0 +1,4824 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +in Vertex { + vec2 vTexCoord; + vec2 tex_uv; + vec2 bloom_dxdy; + float bloom_sigma_runtime; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define MASKED_SCANLINEStexture source[0] +#define MASKED_SCANLINEStexture_size sourceSize[0].xy +#define MASKED_SCANLINESvideo_size sourceSize[0].xy +#define BLOOM_APPROXtexture source[3] +#define BLOOM_APPROXtexture_size sourceSize[3].xy +#define BLOOM_APPROXvideo_size sourceSize[3].xy + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +float bloom_approx_scale_x = targetSize.y / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); +const float bloom_diff_thresh_ = 1.0/256.0; + +////////////////////////////// FRAGMENT INCLUDES ////////////////////////////// + +//#include "bloom-functions.h" + +//////////////////////////// BEGIN BLOOM-FUNCTIONS /////////////////////////// + +#ifndef BLOOM_FUNCTIONS_H +#define BLOOM_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These utility functions and constants help several passes determine the +// size and center texel weight of the phosphor bloom in a uniform manner. + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// We need to calculate the correct blur sigma using some .cgp constants: +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// IN.output_size < IN.video_size. +// 4.) IN.output_size == IN.video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (IN.video_size/IN.output_size)/IN.texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(IN.video_size/IN.output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, IN.output_size.x), [0, IN.output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (IN.position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, IN.output_size.x), [0, IN.output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(IN.video_size/IN.output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +/////////////////////////////// BLOOM CONSTANTS ////////////////////////////// + +// Compute constants with manual inlines of the functions below: +static const float bloom_diff_thresh = 1.0/256.0; + + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +inline float get_absolute_scale_blur_sigma(const float thresh) +{ + // Requires: 1.) min_expected_triads must be a global float. The number + // of horizontal phosphor triads in the final image must be + // >= min_allowed_viewport_triads.x for realistic results. + // 2.) bloom_approx_scale_x must be a global float equal to the + // absolute horizontal scale of BLOOM_APPROX. + // 3.) bloom_approx_scale_x/min_allowed_viewport_triads.x + // should be <= 1.1658025090 to keep the final result < + // 0.62666015625 (the largest sigma ensuring the largest + // unused texel weight stays < 1.0/256.0 for a 3x3 blur). + // 4.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum Gaussian sigma that will blur the pass + // output as much as it would have taken to blur away + // bloom_approx_scale_x horizontal phosphor triads. + // Description: + // BLOOM_APPROX should look like a downscaled phosphor blur. Ideally, we'd + // use the same blur sigma as the actual phosphor bloom and scale it down + // to the current resolution with (bloom_approx_scale_x/viewport_size_x), but + // we don't know the viewport size in this pass. Instead, we'll blur as + // much as it would take to blur away min_allowed_viewport_triads.x. This + // will blur "more than necessary" if the user actually uses more triads, + // but that's not terrible either, because blurring a constant fraction of + // the viewport may better resemble a true optical bloom anyway (since the + // viewport will generally be about the same fraction of each player's + // field of view, regardless of screen size and resolution). + // Assume an extremely large viewport size for asymptotic results. + return bloom_approx_scale_x/max_viewport_size_x * + get_min_sigma_to_blur_triad( + max_viewport_size_x/min_allowed_viewport_triads.x, thresh); +} + +inline float get_center_weight(const float sigma) +{ + // Given a Gaussian blur sigma, get the blur weight for the center texel. + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return get_fast_gaussian_weight_sum_inv(sigma); + #else + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + // Note: If the implementation uses a smaller blur than the max allowed, + // the worst case scenario is that the center weight will be overestimated, + // so we'll put a bit more energy into the brightpass...no huge deal. + // Then again, if the implementation uses a larger blur than the max + // "allowed" because of dynamic branching, the center weight could be + // underestimated, which is more of a problem...consider always using + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // 43x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + + w11 + w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + // 31x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + + w8 + w9 + w10 + w11 + w12 + w13 + w14 + w15)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + // 25x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + // 17x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + #else + // 9x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + const float center_weight = weight_sum_inv * weight_sum_inv; + return center_weight; + #endif +} + +inline float3 tex2DblurNfast(const sampler2D texture, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // If sigma is static, we can safely branch and use the smallest blur + // that's big enough. Ignore #define hints, because we'll only use a + // large blur if we actually need it, and the branches cost nothing. + #ifndef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #else + // It's still worth branching if the profile supports dynamic branches: + // It's much faster than using a hugely excessive blur, but each branch + // eats ~1% FPS. + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #endif + #endif + // Failed optimization notes: + // I originally created a same-size mipmapped 5-tap separable blur10 that + // could handle any sigma by reaching into lower mip levels. It was + // as fast as blur25fast for runtime sigmas and a tad faster than + // blur31fast for static sigmas, but mipmapping two viewport-size passes + // ate 10% of FPS across all codepaths, so it wasn't worth it. + #ifdef PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + if(sigma <= blur9_std_dev) + { + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur17_std_dev) + { + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur25_std_dev) + { + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur31_std_dev) + { + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + } + else + { + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + } + #else + // If we can't afford to branch, we can only guess at what blur + // size we need. Therefore, use the largest blur allowed. + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + #else + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + #endif // PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE +} + +inline float get_bloom_approx_sigma(const float output_size_x_runtime, + const float estimated_viewport_size_x) +{ + // Requires: 1.) output_size_x_runtime == BLOOM_APPROX.output_size.x. + // This is included for dynamic codepaths just in case the + // following two globals are incorrect: + // 2.) bloom_approx_size_x_for_skip should == the same + // if PHOSPHOR_BLOOM_FAKE is #defined + // 3.) bloom_approx_size_x should == the same otherwise + // Returns: For gaussian4x4, return a dynamic small bloom sigma that's + // as close to optimal as possible given available information. + // For blur3x3, return the a static small bloom sigma that + // works well for typical cases. Otherwise, we're using simple + // bilinear filtering, so use static calculations. + // Assume the default static value. This is a compromise that ensures + // typical triads are blurred, even if unusually large ones aren't. + static const float mask_num_triads_static = + max(min_allowed_viewport_triads.x, mask_num_triads_desired_static); + const float mask_num_triads_from_size = + estimated_viewport_size_x/mask_triad_size_desired; + const float mask_num_triads_runtime = max(min_allowed_viewport_triads.x, + lerp(mask_num_triads_from_size, mask_num_triads_desired, + mask_specify_num_triads)); + // Assume an extremely large viewport size for asymptotic results: + static const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + if(bloom_approx_filter > 1.5) // 4x4 true Gaussian resize + { + // Use the runtime num triads and output size: + const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_runtime; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_runtime/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // account for the Gaussian scanline sigma from the last pass too. + // The bloom will be too wide horizontally but tall enough vertically. + return length(float2(bloom_approx_sigma, beam_max_sigma)); + } + else // 3x3 blur resize (the bilinear resize doesn't need a sigma) + { + // We're either using blur3x3 or bilinear filtering. The biggest + // reason to choose blur3x3 is to avoid dynamic weights, so use a + // static calculation. + #ifdef PHOSPHOR_BLOOM_FAKE + static const float output_size_x_static = + bloom_approx_size_x_for_fake; + #else + static const float output_size_x_static = bloom_approx_size_x; + #endif + static const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_static; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_static/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // try accounting for the Gaussian scanline sigma from the last pass + // too; use the static default value: + return length(float2(bloom_approx_sigma, beam_max_sigma_static)); + } +} + +inline float get_final_bloom_sigma(const float bloom_sigma_runtime) +{ + // Requires: 1.) bloom_sigma_runtime is a precalculated sigma that's + // optimal for the [known] triad size. + // 2.) Call this from a fragment shader (not a vertex shader), + // or blurring with static sigmas won't be constant-folded. + // Returns: Return the optimistic static sigma if the triad size is + // known at compile time. Otherwise return the optimal runtime + // sigma (10% slower) or an implementation-specific compromise + // between an optimistic or pessimistic static sigma. + // Notes: Call this from the fragment shader, NOT the vertex shader, + // so static sigmas can be constant-folded! + const float bloom_sigma_optimistic = get_min_sigma_to_blur_triad( + mask_triad_size_desired_static, bloom_diff_thresh); + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return bloom_sigma_runtime; + #else + // Overblurring looks as bad as underblurring, so assume average-size + // triads, not worst-case huge triads: + return bloom_sigma_optimistic; + #endif +} + + +#endif // BLOOM_FUNCTIONS_H + +//////////////////////////// END BLOOM-FUNCTIONS /////////////////////////// + +/////////////////////////// END FRAGMENT-INCLUDES ////////////////////////// + +void main() { + // Blur the brightpass horizontally with a 9/17/25/43x blur: + const float bloom_sigma = get_final_bloom_sigma(bloom_sigma_runtime); + const float3 color = tex2DblurNfast(input_texture, tex_uv, + bloom_dxdy, bloom_sigma); + // Encode and output the blurred image: + FragColor = encode_output(float4(color, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/bloom-vertical.vs b/shaders/CRT-Royale.shader/bloom-vertical.vs new file mode 100644 index 000000000..dfec96e63 --- /dev/null +++ b/shaders/CRT-Royale.shader/bloom-vertical.vs @@ -0,0 +1,3792 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 tex_uv; + vec2 bloom_dxdy; + float bloom_sigma_runtime; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define MASKED_SCANLINEStexture source[0] +#define MASKED_SCANLINEStexture_size sourceSize[0].xy +#define MASKED_SCANLINESvideo_size sourceSize[0].xy +#define BLOOM_APPROXtexture source[3] +#define BLOOM_APPROXtexture_size sourceSize[3].xy +#define BLOOM_APPROXvideo_size sourceSize[3].xy + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-params.h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == IN.output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +///////////////////////////// END VERTEX-INCLUDES //////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +float bloom_approx_scale_x = targetSize.y / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); +const float bloom_diff_thresh_ = 1.0/256.0; + +// copied from bloom-functions.h +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +void main() { + gl_Position = position; + vTexCoord = texCoord; + tex_uv = vTexCoord.xy * 1.0001; + + // Get the uv sample distance between output pixels. Calculate dxdy like + // blurs/vertex-shader-blur-fast-vertical.h. + const float2 dxdy_scale = video_size/output_size; + const float2 dxdy = dxdy_scale/texture_size; + // This blur is vertical-only, so zero out the vertical offset: + bloom_dxdy = float2(0.0, dxdy.y); + + // Calculate a runtime bloom_sigma in case it's needed: + const float mask_tile_size_x = get_resized_mask_tile_size( + output_size, output_size * mask_resize_viewport_scale, false).x; + bloom_sigma_runtime = get_min_sigma_to_blur_triad( + mask_tile_size_x / mask_triads_per_tile, bloom_diff_thresh_); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/blur9fast-horizontal.fs b/shaders/CRT-Royale.shader/blur9fast-horizontal.fs new file mode 100644 index 000000000..c7293eed2 --- /dev/null +++ b/shaders/CRT-Royale.shader/blur9fast-horizontal.fs @@ -0,0 +1,2016 @@ +#version 150 + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +#if __VERSION__ >= 130 +#define COMPAT_TEXTURE texture +#else +#define COMPAT_TEXTURE texture2D +#endif + +#ifdef GL_ES +#define COMPAT_PRECISION mediump +#else +#define COMPAT_PRECISION +#endif + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +in Vertex { + vec2 vTexCoord; + vec2 blur_dxdy; +}; + +out vec4 FragColor; + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// PASS SETTINGS: +// gamma-management.h needs to know what kind of pipeline we're using and +// what pass this is in that pipeline. This will become obsolete if/when we +// can #define things like this in the .cgp preset file. +//#define GAMMA_ENCODE_EVERY_FBO +//#define FIRST_PASS +//#define LAST_PASS +//#define SIMULATE_CRT_ON_LCD +//#define SIMULATE_GBA_ON_LCD +//#define SIMULATE_LCD_ON_CRT +//#define SIMULATE_GBA_ON_CRT + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + float lcd_reference_gamma = 2.5; // To match CRT + float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + float get_crt_gamma() { return crt_gamma; } + float get_gba_gamma() { return gba_gamma; } + float get_lcd_gamma() { return lcd_gamma; } +#else + float get_crt_gamma() { return crt_reference_gamma_high; } + float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + float get_intermediate_gamma() { return intermediate_gamma; } + float get_input_gamma() { return input_gamma; } + float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + float get_input_gamma() { return get_crt_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + float get_input_gamma() { return get_lcd_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + float get_input_gamma() { return ntsc_gamma; } + float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + bool linearize_input = true; + float get_pass_input_gamma() { return get_input_gamma(); } + #else + bool linearize_input = false; + float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + bool gamma_encode_output = true; + float get_pass_output_gamma() { return get_output_gamma(); } + #else + bool gamma_encode_output = false; + float get_pass_output_gamma() { return 1.0; } + #endif +#else + bool linearize_input = true; + bool gamma_encode_output = true; + #ifdef FIRST_PASS + float get_pass_input_gamma() { return get_input_gamma(); } + #else + float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + float get_pass_output_gamma() { return get_output_gamma(); } + #else + float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +vec4 decode_input(vec4 color) +{ + if(linearize_input = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +vec4 encode_output(vec4 color) +{ + if(gamma_encode_output = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords))); } + +//#define tex2D_linearize(C, D, E) decode_input(vec4(COMPAT_TEXTURE(C, D, E))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords, int texel_off) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords, texel_off))); } + +#endif // GAMMA_MANAGEMENT_H + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// IN.output_size < IN.video_size. +// 4.) IN.output_size == IN.video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (IN.video_size/IN.output_size)/IN.texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = vec2(dxdy.x, 0.0) or vec2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(IN.video_size/IN.output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static float blur3_std_dev +// static float blur4_std_dev +// static float blur5_std_dev +// static float blur6_std_dev +// static float blur7_std_dev +// static float blur8_std_dev +// static float blur9_std_dev +// static float blur10_std_dev +// static float blur11_std_dev +// static float blur12_std_dev +// static float blur17_std_dev +// static float blur25_std_dev +// static float blur31_std_dev +// static float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + float blur3_std_dev = 0.84931640625; + float blur4_std_dev = 0.84931640625; + float blur5_std_dev = 1.0595703125; + float blur6_std_dev = 1.06591796875; + float blur7_std_dev = 1.17041015625; + float blur8_std_dev = 1.1720703125; + float blur9_std_dev = 1.2259765625; + float blur10_std_dev = 1.21982421875; + float blur11_std_dev = 1.25361328125; + float blur12_std_dev = 1.2423828125; + float blur17_std_dev = 1.27783203125; + float blur25_std_dev = 1.2810546875; + float blur31_std_dev = 1.28125; + float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + float blur3_std_dev = 0.62666015625; + float blur4_std_dev = 0.66171875; + float blur5_std_dev = 0.9845703125; + float blur6_std_dev = 1.02626953125; + float blur7_std_dev = 1.36103515625; + float blur8_std_dev = 1.4080078125; + float blur9_std_dev = 1.7533203125; + float blur10_std_dev = 1.80478515625; + float blur11_std_dev = 2.15986328125; + float blur12_std_dev = 2.215234375; + float blur17_std_dev = 3.45535583496; + float blur25_std_dev = 5.3409576416; + float blur31_std_dev = 6.86488037109; + float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + float error_blurring = 0.5; +#endif + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" +//#include "quad-pixel-communication.h" +//#include "special-functions.h" + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (vec4/vec3/vec2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +vec4 erf6(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + vec4 one = vec4(1.0); + vec4 sign_x = sign(x); + vec4 t = one/(one + 0.47047*abs(x)); + vec4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec3 erf6(vec3 x) +{ + // vec3 version: + vec3 one = vec3(1.0); + vec3 sign_x = sign(x); + vec3 t = one/(one + 0.47047*abs(x)); + vec3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec2 erf6(vec2 x) +{ + // vec2 version: + vec2 one = vec2(1.0); + vec2 sign_x = sign(x); + vec2 t = one/(one + 0.47047*abs(x)); + vec2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(float x) +{ + // Float version: + float sign_x = sign(x); + float t = 1.0/(1.0 + 0.47047*abs(x)); + float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec4 erft(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +vec3 erft(vec3 x) +{ + // vec3 version: + return tanh(1.202760580 * x); +} + +vec2 erft(vec2 x) +{ + // vec2 version: + return tanh(1.202760580 * x); +} + +float erft(float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +vec4 erf(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec3 erf(vec3 x) +{ + // vec3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec2 erf(vec2 x) +{ + // vec2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +float erf(float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +vec4 gamma_impl(vec4 s, vec4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + vec4 g = vec4(1.12906830989); + vec4 c0 = vec4(0.8109119309638332633713423362694399653724431); + vec4 c1 = vec4(0.4808354605142681877121661197951496120000040); + vec4 e = vec4(2.71828182845904523536028747135266249775724709); + vec4 sph = s + vec4(0.5); + vec4 lanczos_sum = c0 + c1/(s + vec4(1.0)); + vec4 base = (sph + g)/e; // or (s + g + vec4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec3 gamma_impl(vec3 s, vec3 s_inv) +{ + // vec3 version: + vec3 g = vec3(1.12906830989); + vec3 c0 = vec3(0.8109119309638332633713423362694399653724431); + vec3 c1 = vec3(0.4808354605142681877121661197951496120000040); + vec3 e = vec3(2.71828182845904523536028747135266249775724709); + vec3 sph = s + vec3(0.5); + vec3 lanczos_sum = c0 + c1/(s + vec3(1.0)); + vec3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec2 gamma_impl(vec2 s, vec2 s_inv) +{ + // vec2 version: + vec2 g = vec2(1.12906830989); + vec2 c0 = vec2(0.8109119309638332633713423362694399653724431); + vec2 c1 = vec2(0.4808354605142681877121661197951496120000040); + vec2 e = vec2(2.71828182845904523536028747135266249775724709); + vec2 sph = s + vec2(0.5); + vec2 lanczos_sum = c0 + c1/(s + vec2(1.0)); + vec2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(float s, float s_inv) +{ + // Float version: + float g = 1.12906830989; + float c0 = 0.8109119309638332633713423362694399653724431; + float c1 = 0.4808354605142681877121661197951496120000040; + float e = 2.71828182845904523536028747135266249775724709; + float sph = s + 0.5; + float lanczos_sum = c0 + c1/(s + 1.0); + float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec4 gamma(vec4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, vec4(1.0)/s); +} + +vec3 gamma(vec3 s) +{ + // vec3 version: + return gamma_impl(s, vec3(1.0)/s); +} + +vec2 gamma(vec2 s) +{ + // vec2 version: + return gamma_impl(s, vec2(1.0)/s); +} + +float gamma(float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +vec4 ligamma_small_z_impl(vec4 s, vec4 z, vec4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + vec4 scale = pow(z, s); + vec4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + vec4 z_sq = z*z; + vec4 denom1 = s + vec4(1.0); + vec4 denom2 = 2.0*s + vec4(4.0); + vec4 denom3 = 6.0*s + vec4(18.0); + //vec4 denom4 = 24.0*s + vec4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +vec3 ligamma_small_z_impl(vec3 s, vec3 z, vec3 s_inv) +{ + // vec3 version: + vec3 scale = pow(z, s); + vec3 sum = s_inv; + vec3 z_sq = z*z; + vec3 denom1 = s + vec3(1.0); + vec3 denom2 = 2.0*s + vec3(4.0); + vec3 denom3 = 6.0*s + vec3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +vec2 ligamma_small_z_impl(vec2 s, vec2 z, vec2 s_inv) +{ + // vec2 version: + vec2 scale = pow(z, s); + vec2 sum = s_inv; + vec2 z_sq = z*z; + vec2 denom1 = s + vec2(1.0); + vec2 denom2 = 2.0*s + vec2(4.0); + vec2 denom3 = 6.0*s + vec2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(float s, float z, float s_inv) +{ + // Float version: + float scale = pow(z, s); + float sum = s_inv; + float z_sq = z*z; + float denom1 = s + 1.0; + float denom2 = 2.0*s + 4.0; + float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +vec4 uigamma_large_z_impl(vec4 s, vec4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = vec4('inf'); + // vec4 one = vec4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + vec4 numerator = pow(z, s) * exp(-z); + vec4 denom = vec4(7.0) + z - s; + denom = vec4(5.0) + z - s + (3.0*s - vec4(9.0))/denom; + denom = vec4(3.0) + z - s + (2.0*s - vec4(4.0))/denom; + denom = vec4(1.0) + z - s + (s - vec4(1.0))/denom; + return numerator / denom; +} + +vec3 uigamma_large_z_impl(vec3 s, vec3 z) +{ + // vec3 version: + vec3 numerator = pow(z, s) * exp(-z); + vec3 denom = vec3(7.0) + z - s; + denom = vec3(5.0) + z - s + (3.0*s - vec3(9.0))/denom; + denom = vec3(3.0) + z - s + (2.0*s - vec3(4.0))/denom; + denom = vec3(1.0) + z - s + (s - vec3(1.0))/denom; + return numerator / denom; +} + +vec2 uigamma_large_z_impl(vec2 s, vec2 z) +{ + // vec2 version: + vec2 numerator = pow(z, s) * exp(-z); + vec2 denom = vec2(7.0) + z - s; + denom = vec2(5.0) + z - s + (3.0*s - vec2(9.0))/denom; + denom = vec2(3.0) + z - s + (2.0*s - vec2(4.0))/denom; + denom = vec2(1.0) + z - s + (s - vec2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(float s, float z) +{ + // Float version: + float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +vec4 normalized_ligamma_impl(vec4 s, vec4 z, + vec4 s_inv, vec4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + vec4 thresh = vec4(0.775075); + bvec4 z_is_large = greaterThan(z , thresh); + vec4 z_size_check = vec4(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0, z_is_large.w ? 1.0 : 0.0); + vec4 large_z = vec4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + return large_z * vec4(z_size_check) + small_z * vec4(z_size_check); +} + +vec3 normalized_ligamma_impl(vec3 s, vec3 z, + vec3 s_inv, vec3 gamma_s_inv) +{ + // vec3 version: + vec3 thresh = vec3(0.775075); + bvec3 z_is_large = greaterThan(z , thresh); + vec3 z_size_check = vec3(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0); + vec3 large_z = vec3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec3(z_size_check) + small_z * vec3(z_size_check); +} + +vec2 normalized_ligamma_impl(vec2 s, vec2 z, + vec2 s_inv, vec2 gamma_s_inv) +{ + // vec2 version: + vec2 thresh = vec2(0.775075); + bvec2 z_is_large = greaterThan(z , thresh); + vec2 z_size_check = vec2(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0); + vec2 large_z = vec2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec2(z_size_check) + small_z * vec2(z_size_check); +} + +float normalized_ligamma_impl(float s, float z, + float s_inv, float gamma_s_inv) +{ + // Float version: + float thresh = 0.775075; + float z_size_check = 0.0; + if (z > thresh) z_size_check = 1.0; + float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_size_check) + small_z * float(z_size_check); +} + +// Normalized lower incomplete gamma function for small s: +vec4 normalized_ligamma(vec4 s, vec4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + vec4 s_inv = vec4(1.0)/s; + vec4 gamma_s_inv = vec4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec3 normalized_ligamma(vec3 s, vec3 z) +{ + // vec3 version: + vec3 s_inv = vec3(1.0)/s; + vec3 gamma_s_inv = vec3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec2 normalized_ligamma(vec2 s, vec2 z) +{ + // vec2 version: + vec2 s_inv = vec2(1.0)/s; + vec2 gamma_s_inv = vec2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(float s, float z) +{ + // Float version: + float s_inv = 1.0/s; + float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +/////////////////////////////////// HELPERS ////////////////////////////////// + +vec4 uv2_to_uv4(vec2 tex_uv) +{ + // Make a vec2 uv offset safe for adding to vec4 tex2Dlod coords: + return vec4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +float get_fast_gaussian_weight_sum_inv(float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +vec3 tex2Dblur11resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + vec3 sum = vec3(0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +vec3 tex2Dblur11fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w45 = w4 + w5; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur17fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur25fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w9_10 = w9 + w10; + float w11_12 = w11 + w12; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + float w9_10_ratio = w10/w9_10; + float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur31fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur43fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + float w16 = exp(-256.0 * denom_inv); + float w17 = exp(-289.0 * denom_inv); + float w18 = exp(-324.0 * denom_inv); + float w19 = exp(-361.0 * denom_inv); + float w20 = exp(-400.0 * denom_inv); + float w21 = exp(-441.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w16_17 = w16 + w17; + float w18_19 = w18 + w19; + float w20_21 = w20 + w21; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + float w16_17_ratio = w17/w16_17; + float w18_19_ratio = w19/w18_19; + float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + +vec3 tex2Dblur5fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +vec3 tex2Dblur3x3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + vec2 sample4_uv = tex_uv; + vec2 dx = vec2(dxdy.x, 0.0); + vec2 dy = vec2(0.0, dxdy.y); + vec2 sample1_uv = sample4_uv - dy; + vec2 sample7_uv = sample4_uv + dy; + vec3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + vec3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + vec3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + vec3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + vec3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + vec3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + vec3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + vec3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + vec3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + float w4 = 1.0; + float w1_3_5_7 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w0_2_6_8 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + vec3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + +// Resizable one-pass blurs: +vec3 tex2Dblur3x3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w34 = w3 + w4; + float w12_ratio = w2/w12; + float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9x9(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float w4off = exp(-16.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2R_texel_offset = vec2(3.0, 0.0) + vec2(texel3to4ratio, 0.0); + vec2 sample3d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + vec2 sample4d_texel_offset = vec2(3.0, 1.0) + vec2(texel3to4ratio, texel1to2ratio); + vec2 sample5d_texel_offset = vec2(1.0, 3.0) + vec2(texel1to2ratio, texel3to4ratio); + vec2 sample6d_texel_offset = vec2(3.0, 3.0) + vec2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2R1 = w3off; + float w2R2 = w4off; + float w3d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w3d2_3d3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w3d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d1_5d1 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d2_5d3 = exp(-LENGTH_SQ(vec2(4.0, 1.0)) * denom_inv); + float w4d3_5d2 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4_5d4 = exp(-LENGTH_SQ(vec2(4.0, 2.0)) * denom_inv); + float w6d1 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + float w6d2_6d3 = exp(-LENGTH_SQ(vec2(4.0, 3.0)) * denom_inv); + float w6d4 = exp(-LENGTH_SQ(vec2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2R1 + w2R2; + float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + float w5 = w4; + float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + vec3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + vec3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + vec3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + vec3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + vec3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + vec3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + vec3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + vec3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + vec3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7x7(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample1d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + vec2 sample2d_texel_offset = vec2(2.0, 0.0) + vec2(texel2to3ratio, texel0to1ratio); + vec2 sample3d_texel_offset = vec2(0.0, 2.0) + vec2(texel0to1ratio, texel2to3ratio); + vec2 sample4d_texel_offset = vec2(2.0, 2.0) + vec2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1abcd = 1.0; + float w1bd2_1cd3 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w2bd1_3cd1 = exp(-LENGTH_SQ(vec2(2.0, 0.0)) * denom_inv); + float w2bd2_3cd2 = exp(-LENGTH_SQ(vec2(3.0, 0.0)) * denom_inv); + float w1d4 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d3_3d2 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4_3d4 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d1 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d2_4d3 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = vec3(0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5x5(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d2_3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3x3(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample0d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + vec3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + vec3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + vec3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur17fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur17fast(texture, tex_uv, dxdy, blur17_std_dev); +} + +vec3 tex2Dblur25fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur25fast(texture, tex_uv, dxdy, blur25_std_dev); +} + +vec3 tex2Dblur43fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur43fast(texture, tex_uv, dxdy, blur43_std_dev); +} +vec3 tex2Dblur31fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur31fast(texture, tex_uv, dxdy, blur31_std_dev); +} + +vec3 tex2Dblur3fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3fast(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur3x3(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5fast(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur5resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5resize(texture, tex_uv, dxdy, blur5_std_dev); +} +vec3 tex2Dblur3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5x5(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5x5(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur7resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7resize(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7fast(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7x7(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7x7(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur9resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9resize(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur9x9(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9x9(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur11resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11resize(texture, tex_uv, dxdy, blur11_std_dev); +} + +vec3 tex2Dblur11fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11fast(texture, tex_uv, dxdy, blur11_std_dev); +} + +#endif // BLUR_FUNCTIONS_H + +#define Source source[0] +#define tex_uv vTexCoord.xy + +#define InputSize sourceSize[0].xy +#define TextureSize sourceSize[0].xy +#define OutputSize targetSize.xy + +void main() { + vec3 color = tex2Dblur9fast(Source, tex_uv, blur_dxdy); + // Encode and output the blurred image: + FragColor = encode_output(vec4(color, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/blur9fast-horizontal.vs b/shaders/CRT-Royale.shader/blur9fast-horizontal.vs new file mode 100644 index 000000000..7f3b2b942 --- /dev/null +++ b/shaders/CRT-Royale.shader/blur9fast-horizontal.vs @@ -0,0 +1,2025 @@ +#version 150 + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +#if __VERSION__ >= 130 +#define COMPAT_TEXTURE texture +#else +#define COMPAT_TEXTURE texture2D +#endif + +#ifdef GL_ES +#define COMPAT_PRECISION mediump +#else +#define COMPAT_PRECISION +#endif + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 blur_dxdy; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// PASS SETTINGS: +// gamma-management.h needs to know what kind of pipeline we're using and +// what pass this is in that pipeline. This will become obsolete if/when we +// can #define things like this in the .cgp preset file. +//#define GAMMA_ENCODE_EVERY_FBO +//#define FIRST_PASS +//#define LAST_PASS +//#define SIMULATE_CRT_ON_LCD +//#define SIMULATE_GBA_ON_LCD +//#define SIMULATE_LCD_ON_CRT +//#define SIMULATE_GBA_ON_CRT + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + float lcd_reference_gamma = 2.5; // To match CRT + float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + float get_crt_gamma() { return crt_gamma; } + float get_gba_gamma() { return gba_gamma; } + float get_lcd_gamma() { return lcd_gamma; } +#else + float get_crt_gamma() { return crt_reference_gamma_high; } + float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + float get_intermediate_gamma() { return intermediate_gamma; } + float get_input_gamma() { return input_gamma; } + float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + float get_input_gamma() { return get_crt_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + float get_input_gamma() { return get_lcd_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + float get_input_gamma() { return ntsc_gamma; } + float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + bool linearize_input = true; + float get_pass_input_gamma() { return get_input_gamma(); } + #else + bool linearize_input = false; + float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + bool gamma_encode_output = true; + float get_pass_output_gamma() { return get_output_gamma(); } + #else + bool gamma_encode_output = false; + float get_pass_output_gamma() { return 1.0; } + #endif +#else + bool linearize_input = true; + bool gamma_encode_output = true; + #ifdef FIRST_PASS + float get_pass_input_gamma() { return get_input_gamma(); } + #else + float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + float get_pass_output_gamma() { return get_output_gamma(); } + #else + float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +vec4 decode_input(vec4 color) +{ + if(linearize_input = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +vec4 encode_output(vec4 color) +{ + if(gamma_encode_output = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords))); } + +//#define tex2D_linearize(C, D, E) decode_input(vec4(COMPAT_TEXTURE(C, D, E))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords, int texel_off) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords, texel_off))); } + +#endif // GAMMA_MANAGEMENT_H + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// IN.output_size < IN.video_size. +// 4.) IN.output_size == IN.video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (IN.video_size/IN.output_size)/IN.texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = vec2(dxdy.x, 0.0) or vec2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(IN.video_size/IN.output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static float blur3_std_dev +// static float blur4_std_dev +// static float blur5_std_dev +// static float blur6_std_dev +// static float blur7_std_dev +// static float blur8_std_dev +// static float blur9_std_dev +// static float blur10_std_dev +// static float blur11_std_dev +// static float blur12_std_dev +// static float blur17_std_dev +// static float blur25_std_dev +// static float blur31_std_dev +// static float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + float blur3_std_dev = 0.84931640625; + float blur4_std_dev = 0.84931640625; + float blur5_std_dev = 1.0595703125; + float blur6_std_dev = 1.06591796875; + float blur7_std_dev = 1.17041015625; + float blur8_std_dev = 1.1720703125; + float blur9_std_dev = 1.2259765625; + float blur10_std_dev = 1.21982421875; + float blur11_std_dev = 1.25361328125; + float blur12_std_dev = 1.2423828125; + float blur17_std_dev = 1.27783203125; + float blur25_std_dev = 1.2810546875; + float blur31_std_dev = 1.28125; + float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + float blur3_std_dev = 0.62666015625; + float blur4_std_dev = 0.66171875; + float blur5_std_dev = 0.9845703125; + float blur6_std_dev = 1.02626953125; + float blur7_std_dev = 1.36103515625; + float blur8_std_dev = 1.4080078125; + float blur9_std_dev = 1.7533203125; + float blur10_std_dev = 1.80478515625; + float blur11_std_dev = 2.15986328125; + float blur12_std_dev = 2.215234375; + float blur17_std_dev = 3.45535583496; + float blur25_std_dev = 5.3409576416; + float blur31_std_dev = 6.86488037109; + float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + float error_blurring = 0.5; +#endif + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" +//#include "quad-pixel-communication.h" +//#include "special-functions.h" + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (vec4/vec3/vec2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +vec4 erf6(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + vec4 one = vec4(1.0); + vec4 sign_x = sign(x); + vec4 t = one/(one + 0.47047*abs(x)); + vec4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec3 erf6(vec3 x) +{ + // vec3 version: + vec3 one = vec3(1.0); + vec3 sign_x = sign(x); + vec3 t = one/(one + 0.47047*abs(x)); + vec3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec2 erf6(vec2 x) +{ + // vec2 version: + vec2 one = vec2(1.0); + vec2 sign_x = sign(x); + vec2 t = one/(one + 0.47047*abs(x)); + vec2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(float x) +{ + // Float version: + float sign_x = sign(x); + float t = 1.0/(1.0 + 0.47047*abs(x)); + float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec4 erft(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +vec3 erft(vec3 x) +{ + // vec3 version: + return tanh(1.202760580 * x); +} + +vec2 erft(vec2 x) +{ + // vec2 version: + return tanh(1.202760580 * x); +} + +float erft(float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +vec4 erf(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec3 erf(vec3 x) +{ + // vec3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec2 erf(vec2 x) +{ + // vec2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +float erf(float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +vec4 gamma_impl(vec4 s, vec4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + vec4 g = vec4(1.12906830989); + vec4 c0 = vec4(0.8109119309638332633713423362694399653724431); + vec4 c1 = vec4(0.4808354605142681877121661197951496120000040); + vec4 e = vec4(2.71828182845904523536028747135266249775724709); + vec4 sph = s + vec4(0.5); + vec4 lanczos_sum = c0 + c1/(s + vec4(1.0)); + vec4 base = (sph + g)/e; // or (s + g + vec4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec3 gamma_impl(vec3 s, vec3 s_inv) +{ + // vec3 version: + vec3 g = vec3(1.12906830989); + vec3 c0 = vec3(0.8109119309638332633713423362694399653724431); + vec3 c1 = vec3(0.4808354605142681877121661197951496120000040); + vec3 e = vec3(2.71828182845904523536028747135266249775724709); + vec3 sph = s + vec3(0.5); + vec3 lanczos_sum = c0 + c1/(s + vec3(1.0)); + vec3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec2 gamma_impl(vec2 s, vec2 s_inv) +{ + // vec2 version: + vec2 g = vec2(1.12906830989); + vec2 c0 = vec2(0.8109119309638332633713423362694399653724431); + vec2 c1 = vec2(0.4808354605142681877121661197951496120000040); + vec2 e = vec2(2.71828182845904523536028747135266249775724709); + vec2 sph = s + vec2(0.5); + vec2 lanczos_sum = c0 + c1/(s + vec2(1.0)); + vec2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(float s, float s_inv) +{ + // Float version: + float g = 1.12906830989; + float c0 = 0.8109119309638332633713423362694399653724431; + float c1 = 0.4808354605142681877121661197951496120000040; + float e = 2.71828182845904523536028747135266249775724709; + float sph = s + 0.5; + float lanczos_sum = c0 + c1/(s + 1.0); + float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec4 gamma(vec4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, vec4(1.0)/s); +} + +vec3 gamma(vec3 s) +{ + // vec3 version: + return gamma_impl(s, vec3(1.0)/s); +} + +vec2 gamma(vec2 s) +{ + // vec2 version: + return gamma_impl(s, vec2(1.0)/s); +} + +float gamma(float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +vec4 ligamma_small_z_impl(vec4 s, vec4 z, vec4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + vec4 scale = pow(z, s); + vec4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + vec4 z_sq = z*z; + vec4 denom1 = s + vec4(1.0); + vec4 denom2 = 2.0*s + vec4(4.0); + vec4 denom3 = 6.0*s + vec4(18.0); + //vec4 denom4 = 24.0*s + vec4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +vec3 ligamma_small_z_impl(vec3 s, vec3 z, vec3 s_inv) +{ + // vec3 version: + vec3 scale = pow(z, s); + vec3 sum = s_inv; + vec3 z_sq = z*z; + vec3 denom1 = s + vec3(1.0); + vec3 denom2 = 2.0*s + vec3(4.0); + vec3 denom3 = 6.0*s + vec3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +vec2 ligamma_small_z_impl(vec2 s, vec2 z, vec2 s_inv) +{ + // vec2 version: + vec2 scale = pow(z, s); + vec2 sum = s_inv; + vec2 z_sq = z*z; + vec2 denom1 = s + vec2(1.0); + vec2 denom2 = 2.0*s + vec2(4.0); + vec2 denom3 = 6.0*s + vec2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(float s, float z, float s_inv) +{ + // Float version: + float scale = pow(z, s); + float sum = s_inv; + float z_sq = z*z; + float denom1 = s + 1.0; + float denom2 = 2.0*s + 4.0; + float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +vec4 uigamma_large_z_impl(vec4 s, vec4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = vec4('inf'); + // vec4 one = vec4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + vec4 numerator = pow(z, s) * exp(-z); + vec4 denom = vec4(7.0) + z - s; + denom = vec4(5.0) + z - s + (3.0*s - vec4(9.0))/denom; + denom = vec4(3.0) + z - s + (2.0*s - vec4(4.0))/denom; + denom = vec4(1.0) + z - s + (s - vec4(1.0))/denom; + return numerator / denom; +} + +vec3 uigamma_large_z_impl(vec3 s, vec3 z) +{ + // vec3 version: + vec3 numerator = pow(z, s) * exp(-z); + vec3 denom = vec3(7.0) + z - s; + denom = vec3(5.0) + z - s + (3.0*s - vec3(9.0))/denom; + denom = vec3(3.0) + z - s + (2.0*s - vec3(4.0))/denom; + denom = vec3(1.0) + z - s + (s - vec3(1.0))/denom; + return numerator / denom; +} + +vec2 uigamma_large_z_impl(vec2 s, vec2 z) +{ + // vec2 version: + vec2 numerator = pow(z, s) * exp(-z); + vec2 denom = vec2(7.0) + z - s; + denom = vec2(5.0) + z - s + (3.0*s - vec2(9.0))/denom; + denom = vec2(3.0) + z - s + (2.0*s - vec2(4.0))/denom; + denom = vec2(1.0) + z - s + (s - vec2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(float s, float z) +{ + // Float version: + float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +vec4 normalized_ligamma_impl(vec4 s, vec4 z, + vec4 s_inv, vec4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + vec4 thresh = vec4(0.775075); + bvec4 z_is_large = greaterThan(z , thresh); + vec4 z_size_check = vec4(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0, z_is_large.w ? 1.0 : 0.0); + vec4 large_z = vec4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + return large_z * vec4(z_size_check) + small_z * vec4(z_size_check); +} + +vec3 normalized_ligamma_impl(vec3 s, vec3 z, + vec3 s_inv, vec3 gamma_s_inv) +{ + // vec3 version: + vec3 thresh = vec3(0.775075); + bvec3 z_is_large = greaterThan(z , thresh); + vec3 z_size_check = vec3(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0); + vec3 large_z = vec3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec3(z_size_check) + small_z * vec3(z_size_check); +} + +vec2 normalized_ligamma_impl(vec2 s, vec2 z, + vec2 s_inv, vec2 gamma_s_inv) +{ + // vec2 version: + vec2 thresh = vec2(0.775075); + bvec2 z_is_large = greaterThan(z , thresh); + vec2 z_size_check = vec2(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0); + vec2 large_z = vec2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec2(z_size_check) + small_z * vec2(z_size_check); +} + +float normalized_ligamma_impl(float s, float z, + float s_inv, float gamma_s_inv) +{ + // Float version: + float thresh = 0.775075; + float z_size_check = 0.0; + if (z > thresh) z_size_check = 1.0; + float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_size_check) + small_z * float(z_size_check); +} + +// Normalized lower incomplete gamma function for small s: +vec4 normalized_ligamma(vec4 s, vec4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + vec4 s_inv = vec4(1.0)/s; + vec4 gamma_s_inv = vec4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec3 normalized_ligamma(vec3 s, vec3 z) +{ + // vec3 version: + vec3 s_inv = vec3(1.0)/s; + vec3 gamma_s_inv = vec3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec2 normalized_ligamma(vec2 s, vec2 z) +{ + // vec2 version: + vec2 s_inv = vec2(1.0)/s; + vec2 gamma_s_inv = vec2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(float s, float z) +{ + // Float version: + float s_inv = 1.0/s; + float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +/////////////////////////////////// HELPERS ////////////////////////////////// + +vec4 uv2_to_uv4(vec2 tex_uv) +{ + // Make a vec2 uv offset safe for adding to vec4 tex2Dlod coords: + return vec4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +float get_fast_gaussian_weight_sum_inv(float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +vec3 tex2Dblur11resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + vec3 sum = vec3(0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +vec3 tex2Dblur11fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w45 = w4 + w5; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur17fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur25fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w9_10 = w9 + w10; + float w11_12 = w11 + w12; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + float w9_10_ratio = w10/w9_10; + float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur31fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur43fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + float w16 = exp(-256.0 * denom_inv); + float w17 = exp(-289.0 * denom_inv); + float w18 = exp(-324.0 * denom_inv); + float w19 = exp(-361.0 * denom_inv); + float w20 = exp(-400.0 * denom_inv); + float w21 = exp(-441.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w16_17 = w16 + w17; + float w18_19 = w18 + w19; + float w20_21 = w20 + w21; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + float w16_17_ratio = w17/w16_17; + float w18_19_ratio = w19/w18_19; + float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + +vec3 tex2Dblur5fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +vec3 tex2Dblur3x3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + vec2 sample4_uv = tex_uv; + vec2 dx = vec2(dxdy.x, 0.0); + vec2 dy = vec2(0.0, dxdy.y); + vec2 sample1_uv = sample4_uv - dy; + vec2 sample7_uv = sample4_uv + dy; + vec3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + vec3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + vec3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + vec3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + vec3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + vec3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + vec3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + vec3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + vec3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + float w4 = 1.0; + float w1_3_5_7 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w0_2_6_8 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + vec3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + +// Resizable one-pass blurs: +vec3 tex2Dblur3x3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w34 = w3 + w4; + float w12_ratio = w2/w12; + float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9x9(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float w4off = exp(-16.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2R_texel_offset = vec2(3.0, 0.0) + vec2(texel3to4ratio, 0.0); + vec2 sample3d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + vec2 sample4d_texel_offset = vec2(3.0, 1.0) + vec2(texel3to4ratio, texel1to2ratio); + vec2 sample5d_texel_offset = vec2(1.0, 3.0) + vec2(texel1to2ratio, texel3to4ratio); + vec2 sample6d_texel_offset = vec2(3.0, 3.0) + vec2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2R1 = w3off; + float w2R2 = w4off; + float w3d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w3d2_3d3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w3d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d1_5d1 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d2_5d3 = exp(-LENGTH_SQ(vec2(4.0, 1.0)) * denom_inv); + float w4d3_5d2 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4_5d4 = exp(-LENGTH_SQ(vec2(4.0, 2.0)) * denom_inv); + float w6d1 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + float w6d2_6d3 = exp(-LENGTH_SQ(vec2(4.0, 3.0)) * denom_inv); + float w6d4 = exp(-LENGTH_SQ(vec2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2R1 + w2R2; + float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + float w5 = w4; + float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + vec3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + vec3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + vec3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + vec3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + vec3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + vec3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + vec3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + vec3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + vec3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7x7(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample1d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + vec2 sample2d_texel_offset = vec2(2.0, 0.0) + vec2(texel2to3ratio, texel0to1ratio); + vec2 sample3d_texel_offset = vec2(0.0, 2.0) + vec2(texel0to1ratio, texel2to3ratio); + vec2 sample4d_texel_offset = vec2(2.0, 2.0) + vec2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1abcd = 1.0; + float w1bd2_1cd3 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w2bd1_3cd1 = exp(-LENGTH_SQ(vec2(2.0, 0.0)) * denom_inv); + float w2bd2_3cd2 = exp(-LENGTH_SQ(vec2(3.0, 0.0)) * denom_inv); + float w1d4 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d3_3d2 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4_3d4 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d1 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d2_4d3 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = vec3(0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5x5(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d2_3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3x3(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample0d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + vec3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + vec3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + vec3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur17fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur17fast(texture, tex_uv, dxdy, blur17_std_dev); +} + +vec3 tex2Dblur25fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur25fast(texture, tex_uv, dxdy, blur25_std_dev); +} + +vec3 tex2Dblur43fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur43fast(texture, tex_uv, dxdy, blur43_std_dev); +} +vec3 tex2Dblur31fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur31fast(texture, tex_uv, dxdy, blur31_std_dev); +} + +vec3 tex2Dblur3fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3fast(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur3x3(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5fast(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur5resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5resize(texture, tex_uv, dxdy, blur5_std_dev); +} +vec3 tex2Dblur3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5x5(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5x5(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur7resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7resize(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7fast(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7x7(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7x7(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur9resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9resize(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur9x9(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9x9(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur11resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11resize(texture, tex_uv, dxdy, blur11_std_dev); +} + +vec3 tex2Dblur11fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11fast(texture, tex_uv, dxdy, blur11_std_dev); +} + +#endif // BLUR_FUNCTIONS_H + +#define InputSize sourceSize[0].xy +#define TextureSize sourceSize[0].xy +#define OutputSize targetSize.xy + +void main() { + gl_Position = position; + vTexCoord = texCoord; + // Get the uv sample distance between output pixels. Blurs are not generic + // Gaussian resizers, and correct blurs require: + // 1.) OutputSize == InputSize * 2^m, where m is an integer <= 0. + // 2.) mipmap_inputN = "true" for this pass in the preset if m != 0 + // 3.) filter_linearN = "true" except for 1x scale nearest neighbor blurs + // Gaussian resizers would upsize using the distance between input texels + // (not output pixels), but we avoid this and consistently blur at the + // destination size. Otherwise, combining statically calculated weights + // with bilinear sample exploitation would result in terrible artifacts. + vec2 dxdy_scale = InputSize/OutputSize; + vec2 dxdy = dxdy_scale/TextureSize; + // This blur is vertical-only, so zero out the horizontal offset: + blur_dxdy = vec2(dxdy.x, 0.0); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/blur9fast-vertical.fs b/shaders/CRT-Royale.shader/blur9fast-vertical.fs new file mode 100644 index 000000000..c7293eed2 --- /dev/null +++ b/shaders/CRT-Royale.shader/blur9fast-vertical.fs @@ -0,0 +1,2016 @@ +#version 150 + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +#if __VERSION__ >= 130 +#define COMPAT_TEXTURE texture +#else +#define COMPAT_TEXTURE texture2D +#endif + +#ifdef GL_ES +#define COMPAT_PRECISION mediump +#else +#define COMPAT_PRECISION +#endif + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +in Vertex { + vec2 vTexCoord; + vec2 blur_dxdy; +}; + +out vec4 FragColor; + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// PASS SETTINGS: +// gamma-management.h needs to know what kind of pipeline we're using and +// what pass this is in that pipeline. This will become obsolete if/when we +// can #define things like this in the .cgp preset file. +//#define GAMMA_ENCODE_EVERY_FBO +//#define FIRST_PASS +//#define LAST_PASS +//#define SIMULATE_CRT_ON_LCD +//#define SIMULATE_GBA_ON_LCD +//#define SIMULATE_LCD_ON_CRT +//#define SIMULATE_GBA_ON_CRT + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + float lcd_reference_gamma = 2.5; // To match CRT + float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + float get_crt_gamma() { return crt_gamma; } + float get_gba_gamma() { return gba_gamma; } + float get_lcd_gamma() { return lcd_gamma; } +#else + float get_crt_gamma() { return crt_reference_gamma_high; } + float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + float get_intermediate_gamma() { return intermediate_gamma; } + float get_input_gamma() { return input_gamma; } + float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + float get_input_gamma() { return get_crt_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + float get_input_gamma() { return get_lcd_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + float get_input_gamma() { return ntsc_gamma; } + float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + bool linearize_input = true; + float get_pass_input_gamma() { return get_input_gamma(); } + #else + bool linearize_input = false; + float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + bool gamma_encode_output = true; + float get_pass_output_gamma() { return get_output_gamma(); } + #else + bool gamma_encode_output = false; + float get_pass_output_gamma() { return 1.0; } + #endif +#else + bool linearize_input = true; + bool gamma_encode_output = true; + #ifdef FIRST_PASS + float get_pass_input_gamma() { return get_input_gamma(); } + #else + float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + float get_pass_output_gamma() { return get_output_gamma(); } + #else + float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +vec4 decode_input(vec4 color) +{ + if(linearize_input = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +vec4 encode_output(vec4 color) +{ + if(gamma_encode_output = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords))); } + +//#define tex2D_linearize(C, D, E) decode_input(vec4(COMPAT_TEXTURE(C, D, E))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords, int texel_off) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords, texel_off))); } + +#endif // GAMMA_MANAGEMENT_H + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// IN.output_size < IN.video_size. +// 4.) IN.output_size == IN.video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (IN.video_size/IN.output_size)/IN.texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = vec2(dxdy.x, 0.0) or vec2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(IN.video_size/IN.output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static float blur3_std_dev +// static float blur4_std_dev +// static float blur5_std_dev +// static float blur6_std_dev +// static float blur7_std_dev +// static float blur8_std_dev +// static float blur9_std_dev +// static float blur10_std_dev +// static float blur11_std_dev +// static float blur12_std_dev +// static float blur17_std_dev +// static float blur25_std_dev +// static float blur31_std_dev +// static float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + float blur3_std_dev = 0.84931640625; + float blur4_std_dev = 0.84931640625; + float blur5_std_dev = 1.0595703125; + float blur6_std_dev = 1.06591796875; + float blur7_std_dev = 1.17041015625; + float blur8_std_dev = 1.1720703125; + float blur9_std_dev = 1.2259765625; + float blur10_std_dev = 1.21982421875; + float blur11_std_dev = 1.25361328125; + float blur12_std_dev = 1.2423828125; + float blur17_std_dev = 1.27783203125; + float blur25_std_dev = 1.2810546875; + float blur31_std_dev = 1.28125; + float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + float blur3_std_dev = 0.62666015625; + float blur4_std_dev = 0.66171875; + float blur5_std_dev = 0.9845703125; + float blur6_std_dev = 1.02626953125; + float blur7_std_dev = 1.36103515625; + float blur8_std_dev = 1.4080078125; + float blur9_std_dev = 1.7533203125; + float blur10_std_dev = 1.80478515625; + float blur11_std_dev = 2.15986328125; + float blur12_std_dev = 2.215234375; + float blur17_std_dev = 3.45535583496; + float blur25_std_dev = 5.3409576416; + float blur31_std_dev = 6.86488037109; + float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + float error_blurring = 0.5; +#endif + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" +//#include "quad-pixel-communication.h" +//#include "special-functions.h" + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (vec4/vec3/vec2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +vec4 erf6(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + vec4 one = vec4(1.0); + vec4 sign_x = sign(x); + vec4 t = one/(one + 0.47047*abs(x)); + vec4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec3 erf6(vec3 x) +{ + // vec3 version: + vec3 one = vec3(1.0); + vec3 sign_x = sign(x); + vec3 t = one/(one + 0.47047*abs(x)); + vec3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec2 erf6(vec2 x) +{ + // vec2 version: + vec2 one = vec2(1.0); + vec2 sign_x = sign(x); + vec2 t = one/(one + 0.47047*abs(x)); + vec2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(float x) +{ + // Float version: + float sign_x = sign(x); + float t = 1.0/(1.0 + 0.47047*abs(x)); + float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec4 erft(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +vec3 erft(vec3 x) +{ + // vec3 version: + return tanh(1.202760580 * x); +} + +vec2 erft(vec2 x) +{ + // vec2 version: + return tanh(1.202760580 * x); +} + +float erft(float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +vec4 erf(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec3 erf(vec3 x) +{ + // vec3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec2 erf(vec2 x) +{ + // vec2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +float erf(float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +vec4 gamma_impl(vec4 s, vec4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + vec4 g = vec4(1.12906830989); + vec4 c0 = vec4(0.8109119309638332633713423362694399653724431); + vec4 c1 = vec4(0.4808354605142681877121661197951496120000040); + vec4 e = vec4(2.71828182845904523536028747135266249775724709); + vec4 sph = s + vec4(0.5); + vec4 lanczos_sum = c0 + c1/(s + vec4(1.0)); + vec4 base = (sph + g)/e; // or (s + g + vec4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec3 gamma_impl(vec3 s, vec3 s_inv) +{ + // vec3 version: + vec3 g = vec3(1.12906830989); + vec3 c0 = vec3(0.8109119309638332633713423362694399653724431); + vec3 c1 = vec3(0.4808354605142681877121661197951496120000040); + vec3 e = vec3(2.71828182845904523536028747135266249775724709); + vec3 sph = s + vec3(0.5); + vec3 lanczos_sum = c0 + c1/(s + vec3(1.0)); + vec3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec2 gamma_impl(vec2 s, vec2 s_inv) +{ + // vec2 version: + vec2 g = vec2(1.12906830989); + vec2 c0 = vec2(0.8109119309638332633713423362694399653724431); + vec2 c1 = vec2(0.4808354605142681877121661197951496120000040); + vec2 e = vec2(2.71828182845904523536028747135266249775724709); + vec2 sph = s + vec2(0.5); + vec2 lanczos_sum = c0 + c1/(s + vec2(1.0)); + vec2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(float s, float s_inv) +{ + // Float version: + float g = 1.12906830989; + float c0 = 0.8109119309638332633713423362694399653724431; + float c1 = 0.4808354605142681877121661197951496120000040; + float e = 2.71828182845904523536028747135266249775724709; + float sph = s + 0.5; + float lanczos_sum = c0 + c1/(s + 1.0); + float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec4 gamma(vec4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, vec4(1.0)/s); +} + +vec3 gamma(vec3 s) +{ + // vec3 version: + return gamma_impl(s, vec3(1.0)/s); +} + +vec2 gamma(vec2 s) +{ + // vec2 version: + return gamma_impl(s, vec2(1.0)/s); +} + +float gamma(float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +vec4 ligamma_small_z_impl(vec4 s, vec4 z, vec4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + vec4 scale = pow(z, s); + vec4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + vec4 z_sq = z*z; + vec4 denom1 = s + vec4(1.0); + vec4 denom2 = 2.0*s + vec4(4.0); + vec4 denom3 = 6.0*s + vec4(18.0); + //vec4 denom4 = 24.0*s + vec4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +vec3 ligamma_small_z_impl(vec3 s, vec3 z, vec3 s_inv) +{ + // vec3 version: + vec3 scale = pow(z, s); + vec3 sum = s_inv; + vec3 z_sq = z*z; + vec3 denom1 = s + vec3(1.0); + vec3 denom2 = 2.0*s + vec3(4.0); + vec3 denom3 = 6.0*s + vec3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +vec2 ligamma_small_z_impl(vec2 s, vec2 z, vec2 s_inv) +{ + // vec2 version: + vec2 scale = pow(z, s); + vec2 sum = s_inv; + vec2 z_sq = z*z; + vec2 denom1 = s + vec2(1.0); + vec2 denom2 = 2.0*s + vec2(4.0); + vec2 denom3 = 6.0*s + vec2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(float s, float z, float s_inv) +{ + // Float version: + float scale = pow(z, s); + float sum = s_inv; + float z_sq = z*z; + float denom1 = s + 1.0; + float denom2 = 2.0*s + 4.0; + float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +vec4 uigamma_large_z_impl(vec4 s, vec4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = vec4('inf'); + // vec4 one = vec4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + vec4 numerator = pow(z, s) * exp(-z); + vec4 denom = vec4(7.0) + z - s; + denom = vec4(5.0) + z - s + (3.0*s - vec4(9.0))/denom; + denom = vec4(3.0) + z - s + (2.0*s - vec4(4.0))/denom; + denom = vec4(1.0) + z - s + (s - vec4(1.0))/denom; + return numerator / denom; +} + +vec3 uigamma_large_z_impl(vec3 s, vec3 z) +{ + // vec3 version: + vec3 numerator = pow(z, s) * exp(-z); + vec3 denom = vec3(7.0) + z - s; + denom = vec3(5.0) + z - s + (3.0*s - vec3(9.0))/denom; + denom = vec3(3.0) + z - s + (2.0*s - vec3(4.0))/denom; + denom = vec3(1.0) + z - s + (s - vec3(1.0))/denom; + return numerator / denom; +} + +vec2 uigamma_large_z_impl(vec2 s, vec2 z) +{ + // vec2 version: + vec2 numerator = pow(z, s) * exp(-z); + vec2 denom = vec2(7.0) + z - s; + denom = vec2(5.0) + z - s + (3.0*s - vec2(9.0))/denom; + denom = vec2(3.0) + z - s + (2.0*s - vec2(4.0))/denom; + denom = vec2(1.0) + z - s + (s - vec2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(float s, float z) +{ + // Float version: + float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +vec4 normalized_ligamma_impl(vec4 s, vec4 z, + vec4 s_inv, vec4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + vec4 thresh = vec4(0.775075); + bvec4 z_is_large = greaterThan(z , thresh); + vec4 z_size_check = vec4(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0, z_is_large.w ? 1.0 : 0.0); + vec4 large_z = vec4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + return large_z * vec4(z_size_check) + small_z * vec4(z_size_check); +} + +vec3 normalized_ligamma_impl(vec3 s, vec3 z, + vec3 s_inv, vec3 gamma_s_inv) +{ + // vec3 version: + vec3 thresh = vec3(0.775075); + bvec3 z_is_large = greaterThan(z , thresh); + vec3 z_size_check = vec3(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0); + vec3 large_z = vec3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec3(z_size_check) + small_z * vec3(z_size_check); +} + +vec2 normalized_ligamma_impl(vec2 s, vec2 z, + vec2 s_inv, vec2 gamma_s_inv) +{ + // vec2 version: + vec2 thresh = vec2(0.775075); + bvec2 z_is_large = greaterThan(z , thresh); + vec2 z_size_check = vec2(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0); + vec2 large_z = vec2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec2(z_size_check) + small_z * vec2(z_size_check); +} + +float normalized_ligamma_impl(float s, float z, + float s_inv, float gamma_s_inv) +{ + // Float version: + float thresh = 0.775075; + float z_size_check = 0.0; + if (z > thresh) z_size_check = 1.0; + float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_size_check) + small_z * float(z_size_check); +} + +// Normalized lower incomplete gamma function for small s: +vec4 normalized_ligamma(vec4 s, vec4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + vec4 s_inv = vec4(1.0)/s; + vec4 gamma_s_inv = vec4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec3 normalized_ligamma(vec3 s, vec3 z) +{ + // vec3 version: + vec3 s_inv = vec3(1.0)/s; + vec3 gamma_s_inv = vec3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec2 normalized_ligamma(vec2 s, vec2 z) +{ + // vec2 version: + vec2 s_inv = vec2(1.0)/s; + vec2 gamma_s_inv = vec2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(float s, float z) +{ + // Float version: + float s_inv = 1.0/s; + float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +/////////////////////////////////// HELPERS ////////////////////////////////// + +vec4 uv2_to_uv4(vec2 tex_uv) +{ + // Make a vec2 uv offset safe for adding to vec4 tex2Dlod coords: + return vec4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +float get_fast_gaussian_weight_sum_inv(float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +vec3 tex2Dblur11resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + vec3 sum = vec3(0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +vec3 tex2Dblur11fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w45 = w4 + w5; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur17fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur25fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w9_10 = w9 + w10; + float w11_12 = w11 + w12; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + float w9_10_ratio = w10/w9_10; + float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur31fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur43fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + float w16 = exp(-256.0 * denom_inv); + float w17 = exp(-289.0 * denom_inv); + float w18 = exp(-324.0 * denom_inv); + float w19 = exp(-361.0 * denom_inv); + float w20 = exp(-400.0 * denom_inv); + float w21 = exp(-441.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w16_17 = w16 + w17; + float w18_19 = w18 + w19; + float w20_21 = w20 + w21; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + float w16_17_ratio = w17/w16_17; + float w18_19_ratio = w19/w18_19; + float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + +vec3 tex2Dblur5fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +vec3 tex2Dblur3x3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + vec2 sample4_uv = tex_uv; + vec2 dx = vec2(dxdy.x, 0.0); + vec2 dy = vec2(0.0, dxdy.y); + vec2 sample1_uv = sample4_uv - dy; + vec2 sample7_uv = sample4_uv + dy; + vec3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + vec3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + vec3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + vec3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + vec3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + vec3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + vec3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + vec3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + vec3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + float w4 = 1.0; + float w1_3_5_7 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w0_2_6_8 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + vec3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + +// Resizable one-pass blurs: +vec3 tex2Dblur3x3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w34 = w3 + w4; + float w12_ratio = w2/w12; + float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9x9(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float w4off = exp(-16.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2R_texel_offset = vec2(3.0, 0.0) + vec2(texel3to4ratio, 0.0); + vec2 sample3d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + vec2 sample4d_texel_offset = vec2(3.0, 1.0) + vec2(texel3to4ratio, texel1to2ratio); + vec2 sample5d_texel_offset = vec2(1.0, 3.0) + vec2(texel1to2ratio, texel3to4ratio); + vec2 sample6d_texel_offset = vec2(3.0, 3.0) + vec2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2R1 = w3off; + float w2R2 = w4off; + float w3d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w3d2_3d3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w3d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d1_5d1 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d2_5d3 = exp(-LENGTH_SQ(vec2(4.0, 1.0)) * denom_inv); + float w4d3_5d2 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4_5d4 = exp(-LENGTH_SQ(vec2(4.0, 2.0)) * denom_inv); + float w6d1 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + float w6d2_6d3 = exp(-LENGTH_SQ(vec2(4.0, 3.0)) * denom_inv); + float w6d4 = exp(-LENGTH_SQ(vec2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2R1 + w2R2; + float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + float w5 = w4; + float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + vec3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + vec3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + vec3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + vec3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + vec3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + vec3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + vec3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + vec3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + vec3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7x7(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample1d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + vec2 sample2d_texel_offset = vec2(2.0, 0.0) + vec2(texel2to3ratio, texel0to1ratio); + vec2 sample3d_texel_offset = vec2(0.0, 2.0) + vec2(texel0to1ratio, texel2to3ratio); + vec2 sample4d_texel_offset = vec2(2.0, 2.0) + vec2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1abcd = 1.0; + float w1bd2_1cd3 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w2bd1_3cd1 = exp(-LENGTH_SQ(vec2(2.0, 0.0)) * denom_inv); + float w2bd2_3cd2 = exp(-LENGTH_SQ(vec2(3.0, 0.0)) * denom_inv); + float w1d4 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d3_3d2 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4_3d4 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d1 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d2_4d3 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = vec3(0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5x5(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d2_3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3x3(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample0d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + vec3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + vec3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + vec3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur17fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur17fast(texture, tex_uv, dxdy, blur17_std_dev); +} + +vec3 tex2Dblur25fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur25fast(texture, tex_uv, dxdy, blur25_std_dev); +} + +vec3 tex2Dblur43fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur43fast(texture, tex_uv, dxdy, blur43_std_dev); +} +vec3 tex2Dblur31fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur31fast(texture, tex_uv, dxdy, blur31_std_dev); +} + +vec3 tex2Dblur3fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3fast(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur3x3(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5fast(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur5resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5resize(texture, tex_uv, dxdy, blur5_std_dev); +} +vec3 tex2Dblur3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5x5(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5x5(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur7resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7resize(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7fast(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7x7(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7x7(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur9resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9resize(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur9x9(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9x9(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur11resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11resize(texture, tex_uv, dxdy, blur11_std_dev); +} + +vec3 tex2Dblur11fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11fast(texture, tex_uv, dxdy, blur11_std_dev); +} + +#endif // BLUR_FUNCTIONS_H + +#define Source source[0] +#define tex_uv vTexCoord.xy + +#define InputSize sourceSize[0].xy +#define TextureSize sourceSize[0].xy +#define OutputSize targetSize.xy + +void main() { + vec3 color = tex2Dblur9fast(Source, tex_uv, blur_dxdy); + // Encode and output the blurred image: + FragColor = encode_output(vec4(color, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/blur9fast-vertical.vs b/shaders/CRT-Royale.shader/blur9fast-vertical.vs new file mode 100644 index 000000000..8c10ad960 --- /dev/null +++ b/shaders/CRT-Royale.shader/blur9fast-vertical.vs @@ -0,0 +1,2025 @@ +#version 150 + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +#if __VERSION__ >= 130 +#define COMPAT_TEXTURE texture +#else +#define COMPAT_TEXTURE texture2D +#endif + +#ifdef GL_ES +#define COMPAT_PRECISION mediump +#else +#define COMPAT_PRECISION +#endif + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 blur_dxdy; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// PASS SETTINGS: +// gamma-management.h needs to know what kind of pipeline we're using and +// what pass this is in that pipeline. This will become obsolete if/when we +// can #define things like this in the .cgp preset file. +//#define GAMMA_ENCODE_EVERY_FBO +//#define FIRST_PASS +//#define LAST_PASS +//#define SIMULATE_CRT_ON_LCD +//#define SIMULATE_GBA_ON_LCD +//#define SIMULATE_LCD_ON_CRT +//#define SIMULATE_GBA_ON_CRT + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + float lcd_reference_gamma = 2.5; // To match CRT + float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + float get_crt_gamma() { return crt_gamma; } + float get_gba_gamma() { return gba_gamma; } + float get_lcd_gamma() { return lcd_gamma; } +#else + float get_crt_gamma() { return crt_reference_gamma_high; } + float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + float get_intermediate_gamma() { return intermediate_gamma; } + float get_input_gamma() { return input_gamma; } + float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + float get_input_gamma() { return get_crt_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + float get_input_gamma() { return get_lcd_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + float get_input_gamma() { return get_gba_gamma(); } + float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + float get_input_gamma() { return ntsc_gamma; } + float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + bool linearize_input = true; + float get_pass_input_gamma() { return get_input_gamma(); } + #else + bool linearize_input = false; + float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + bool gamma_encode_output = true; + float get_pass_output_gamma() { return get_output_gamma(); } + #else + bool gamma_encode_output = false; + float get_pass_output_gamma() { return 1.0; } + #endif +#else + bool linearize_input = true; + bool gamma_encode_output = true; + #ifdef FIRST_PASS + float get_pass_input_gamma() { return get_input_gamma(); } + #else + float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + float get_pass_output_gamma() { return get_output_gamma(); } + #else + float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +vec4 decode_input(vec4 color) +{ + if(linearize_input = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +vec4 encode_output(vec4 color) +{ + if(gamma_encode_output = true) + { + if(assume_opaque_alpha = true) + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return vec4(pow(color.rgb, vec3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords))); } + +//#define tex2D_linearize(C, D, E) decode_input(vec4(COMPAT_TEXTURE(C, D, E))) +//vec4 tex2D_linearize(sampler2D tex, vec2 tex_coords, int texel_off) +//{ return decode_input(vec4(COMPAT_TEXTURE(tex, tex_coords, texel_off))); } + +#endif // GAMMA_MANAGEMENT_H + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// IN.output_size < IN.video_size. +// 4.) IN.output_size == IN.video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (IN.video_size/IN.output_size)/IN.texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = vec2(dxdy.x, 0.0) or vec2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(IN.video_size/IN.output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static float blur3_std_dev +// static float blur4_std_dev +// static float blur5_std_dev +// static float blur6_std_dev +// static float blur7_std_dev +// static float blur8_std_dev +// static float blur9_std_dev +// static float blur10_std_dev +// static float blur11_std_dev +// static float blur12_std_dev +// static float blur17_std_dev +// static float blur25_std_dev +// static float blur31_std_dev +// static float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + float blur3_std_dev = 0.84931640625; + float blur4_std_dev = 0.84931640625; + float blur5_std_dev = 1.0595703125; + float blur6_std_dev = 1.06591796875; + float blur7_std_dev = 1.17041015625; + float blur8_std_dev = 1.1720703125; + float blur9_std_dev = 1.2259765625; + float blur10_std_dev = 1.21982421875; + float blur11_std_dev = 1.25361328125; + float blur12_std_dev = 1.2423828125; + float blur17_std_dev = 1.27783203125; + float blur25_std_dev = 1.2810546875; + float blur31_std_dev = 1.28125; + float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + float blur3_std_dev = 0.62666015625; + float blur4_std_dev = 0.66171875; + float blur5_std_dev = 0.9845703125; + float blur6_std_dev = 1.02626953125; + float blur7_std_dev = 1.36103515625; + float blur8_std_dev = 1.4080078125; + float blur9_std_dev = 1.7533203125; + float blur10_std_dev = 1.80478515625; + float blur11_std_dev = 2.15986328125; + float blur12_std_dev = 2.215234375; + float blur17_std_dev = 3.45535583496; + float blur25_std_dev = 5.3409576416; + float blur31_std_dev = 6.86488037109; + float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + float error_blurring = 0.5; +#endif + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" +//#include "quad-pixel-communication.h" +//#include "special-functions.h" + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (vec4/vec3/vec2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +vec4 erf6(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + vec4 one = vec4(1.0); + vec4 sign_x = sign(x); + vec4 t = one/(one + 0.47047*abs(x)); + vec4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec3 erf6(vec3 x) +{ + // vec3 version: + vec3 one = vec3(1.0); + vec3 sign_x = sign(x); + vec3 t = one/(one + 0.47047*abs(x)); + vec3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec2 erf6(vec2 x) +{ + // vec2 version: + vec2 one = vec2(1.0); + vec2 sign_x = sign(x); + vec2 t = one/(one + 0.47047*abs(x)); + vec2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(float x) +{ + // Float version: + float sign_x = sign(x); + float t = 1.0/(1.0 + 0.47047*abs(x)); + float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +vec4 erft(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +vec3 erft(vec3 x) +{ + // vec3 version: + return tanh(1.202760580 * x); +} + +vec2 erft(vec2 x) +{ + // vec2 version: + return tanh(1.202760580 * x); +} + +float erft(float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +vec4 erf(vec4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec3 erf(vec3 x) +{ + // vec3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +vec2 erf(vec2 x) +{ + // vec2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +float erf(float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +vec4 gamma_impl(vec4 s, vec4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + vec4 g = vec4(1.12906830989); + vec4 c0 = vec4(0.8109119309638332633713423362694399653724431); + vec4 c1 = vec4(0.4808354605142681877121661197951496120000040); + vec4 e = vec4(2.71828182845904523536028747135266249775724709); + vec4 sph = s + vec4(0.5); + vec4 lanczos_sum = c0 + c1/(s + vec4(1.0)); + vec4 base = (sph + g)/e; // or (s + g + vec4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec3 gamma_impl(vec3 s, vec3 s_inv) +{ + // vec3 version: + vec3 g = vec3(1.12906830989); + vec3 c0 = vec3(0.8109119309638332633713423362694399653724431); + vec3 c1 = vec3(0.4808354605142681877121661197951496120000040); + vec3 e = vec3(2.71828182845904523536028747135266249775724709); + vec3 sph = s + vec3(0.5); + vec3 lanczos_sum = c0 + c1/(s + vec3(1.0)); + vec3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec2 gamma_impl(vec2 s, vec2 s_inv) +{ + // vec2 version: + vec2 g = vec2(1.12906830989); + vec2 c0 = vec2(0.8109119309638332633713423362694399653724431); + vec2 c1 = vec2(0.4808354605142681877121661197951496120000040); + vec2 e = vec2(2.71828182845904523536028747135266249775724709); + vec2 sph = s + vec2(0.5); + vec2 lanczos_sum = c0 + c1/(s + vec2(1.0)); + vec2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(float s, float s_inv) +{ + // Float version: + float g = 1.12906830989; + float c0 = 0.8109119309638332633713423362694399653724431; + float c1 = 0.4808354605142681877121661197951496120000040; + float e = 2.71828182845904523536028747135266249775724709; + float sph = s + 0.5; + float lanczos_sum = c0 + c1/(s + 1.0); + float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +vec4 gamma(vec4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, vec4(1.0)/s); +} + +vec3 gamma(vec3 s) +{ + // vec3 version: + return gamma_impl(s, vec3(1.0)/s); +} + +vec2 gamma(vec2 s) +{ + // vec2 version: + return gamma_impl(s, vec2(1.0)/s); +} + +float gamma(float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +vec4 ligamma_small_z_impl(vec4 s, vec4 z, vec4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + vec4 scale = pow(z, s); + vec4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + vec4 z_sq = z*z; + vec4 denom1 = s + vec4(1.0); + vec4 denom2 = 2.0*s + vec4(4.0); + vec4 denom3 = 6.0*s + vec4(18.0); + //vec4 denom4 = 24.0*s + vec4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +vec3 ligamma_small_z_impl(vec3 s, vec3 z, vec3 s_inv) +{ + // vec3 version: + vec3 scale = pow(z, s); + vec3 sum = s_inv; + vec3 z_sq = z*z; + vec3 denom1 = s + vec3(1.0); + vec3 denom2 = 2.0*s + vec3(4.0); + vec3 denom3 = 6.0*s + vec3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +vec2 ligamma_small_z_impl(vec2 s, vec2 z, vec2 s_inv) +{ + // vec2 version: + vec2 scale = pow(z, s); + vec2 sum = s_inv; + vec2 z_sq = z*z; + vec2 denom1 = s + vec2(1.0); + vec2 denom2 = 2.0*s + vec2(4.0); + vec2 denom3 = 6.0*s + vec2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(float s, float z, float s_inv) +{ + // Float version: + float scale = pow(z, s); + float sum = s_inv; + float z_sq = z*z; + float denom1 = s + 1.0; + float denom2 = 2.0*s + 4.0; + float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +vec4 uigamma_large_z_impl(vec4 s, vec4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = vec4('inf'); + // vec4 one = vec4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + vec4 numerator = pow(z, s) * exp(-z); + vec4 denom = vec4(7.0) + z - s; + denom = vec4(5.0) + z - s + (3.0*s - vec4(9.0))/denom; + denom = vec4(3.0) + z - s + (2.0*s - vec4(4.0))/denom; + denom = vec4(1.0) + z - s + (s - vec4(1.0))/denom; + return numerator / denom; +} + +vec3 uigamma_large_z_impl(vec3 s, vec3 z) +{ + // vec3 version: + vec3 numerator = pow(z, s) * exp(-z); + vec3 denom = vec3(7.0) + z - s; + denom = vec3(5.0) + z - s + (3.0*s - vec3(9.0))/denom; + denom = vec3(3.0) + z - s + (2.0*s - vec3(4.0))/denom; + denom = vec3(1.0) + z - s + (s - vec3(1.0))/denom; + return numerator / denom; +} + +vec2 uigamma_large_z_impl(vec2 s, vec2 z) +{ + // vec2 version: + vec2 numerator = pow(z, s) * exp(-z); + vec2 denom = vec2(7.0) + z - s; + denom = vec2(5.0) + z - s + (3.0*s - vec2(9.0))/denom; + denom = vec2(3.0) + z - s + (2.0*s - vec2(4.0))/denom; + denom = vec2(1.0) + z - s + (s - vec2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(float s, float z) +{ + // Float version: + float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +vec4 normalized_ligamma_impl(vec4 s, vec4 z, + vec4 s_inv, vec4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + vec4 thresh = vec4(0.775075); + bvec4 z_is_large = greaterThan(z , thresh); + vec4 z_size_check = vec4(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0, z_is_large.w ? 1.0 : 0.0); + vec4 large_z = vec4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + return large_z * vec4(z_size_check) + small_z * vec4(z_size_check); +} + +vec3 normalized_ligamma_impl(vec3 s, vec3 z, + vec3 s_inv, vec3 gamma_s_inv) +{ + // vec3 version: + vec3 thresh = vec3(0.775075); + bvec3 z_is_large = greaterThan(z , thresh); + vec3 z_size_check = vec3(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0, z_is_large.z ? 1.0 : 0.0); + vec3 large_z = vec3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec3(z_size_check) + small_z * vec3(z_size_check); +} + +vec2 normalized_ligamma_impl(vec2 s, vec2 z, + vec2 s_inv, vec2 gamma_s_inv) +{ + // vec2 version: + vec2 thresh = vec2(0.775075); + bvec2 z_is_large = greaterThan(z , thresh); + vec2 z_size_check = vec2(z_is_large.x ? 1.0 : 0.0, z_is_large.y ? 1.0 : 0.0); + vec2 large_z = vec2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + vec2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * vec2(z_size_check) + small_z * vec2(z_size_check); +} + +float normalized_ligamma_impl(float s, float z, + float s_inv, float gamma_s_inv) +{ + // Float version: + float thresh = 0.775075; + float z_size_check = 0.0; + if (z > thresh) z_size_check = 1.0; + float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_size_check) + small_z * float(z_size_check); +} + +// Normalized lower incomplete gamma function for small s: +vec4 normalized_ligamma(vec4 s, vec4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + vec4 s_inv = vec4(1.0)/s; + vec4 gamma_s_inv = vec4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec3 normalized_ligamma(vec3 s, vec3 z) +{ + // vec3 version: + vec3 s_inv = vec3(1.0)/s; + vec3 gamma_s_inv = vec3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +vec2 normalized_ligamma(vec2 s, vec2 z) +{ + // vec2 version: + vec2 s_inv = vec2(1.0)/s; + vec2 gamma_s_inv = vec2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(float s, float z) +{ + // Float version: + float s_inv = 1.0/s; + float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +/////////////////////////////////// HELPERS ////////////////////////////////// + +vec4 uv2_to_uv4(vec2 tex_uv) +{ + // Make a vec2 uv offset safe for adding to vec4 tex2Dlod coords: + return vec4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +float get_fast_gaussian_weight_sum_inv(float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +vec3 tex2Dblur11resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + vec3 sum = vec3(0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +vec3 tex2Dblur11fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w45 = w4 + w5; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur17fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur25fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + //float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + float w1_2 = w1 + w2; + float w3_4 = w3 + w4; + float w5_6 = w5 + w6; + float w7_8 = w7 + w8; + float w9_10 = w9 + w10; + float w11_12 = w11 + w12; + float w1_2_ratio = w2/w1_2; + float w3_4_ratio = w4/w3_4; + float w5_6_ratio = w6/w5_6; + float w7_8_ratio = w8/w7_8; + float w9_10_ratio = w10/w9_10; + float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur31fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur43fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float w5 = exp(-25.0 * denom_inv); + float w6 = exp(-36.0 * denom_inv); + float w7 = exp(-49.0 * denom_inv); + float w8 = exp(-64.0 * denom_inv); + float w9 = exp(-81.0 * denom_inv); + float w10 = exp(-100.0 * denom_inv); + float w11 = exp(-121.0 * denom_inv); + float w12 = exp(-144.0 * denom_inv); + float w13 = exp(-169.0 * denom_inv); + float w14 = exp(-196.0 * denom_inv); + float w15 = exp(-225.0 * denom_inv); + float w16 = exp(-256.0 * denom_inv); + float w17 = exp(-289.0 * denom_inv); + float w18 = exp(-324.0 * denom_inv); + float w19 = exp(-361.0 * denom_inv); + float w20 = exp(-400.0 * denom_inv); + float w21 = exp(-441.0 * denom_inv); + //float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w0_1 = w0 * 0.5 + w1; + float w2_3 = w2 + w3; + float w4_5 = w4 + w5; + float w6_7 = w6 + w7; + float w8_9 = w8 + w9; + float w10_11 = w10 + w11; + float w12_13 = w12 + w13; + float w14_15 = w14 + w15; + float w16_17 = w16 + w17; + float w18_19 = w18 + w19; + float w20_21 = w20 + w21; + float w0_1_ratio = w1/w0_1; + float w2_3_ratio = w3/w2_3; + float w4_5_ratio = w5/w4_5; + float w6_7_ratio = w7/w6_7; + float w8_9_ratio = w9/w8_9; + float w10_11_ratio = w11/w10_11; + float w12_13_ratio = w13/w12_13; + float w14_15_ratio = w15/w14_15; + float w16_17_ratio = w17/w16_17; + float w18_19_ratio = w19/w18_19; + float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + +vec3 tex2Dblur5fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + float w01 = w0 * 0.5 + w1; + float w23 = w2 + w3; + float w01_ratio = w1/w01; + float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +vec3 tex2Dblur3x3resize(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + vec2 sample4_uv = tex_uv; + vec2 dx = vec2(dxdy.x, 0.0); + vec2 dy = vec2(0.0, dxdy.y); + vec2 sample1_uv = sample4_uv - dy; + vec2 sample7_uv = sample4_uv + dy; + vec3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + vec3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + vec3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + vec3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + vec3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + vec3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + vec3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + vec3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + vec3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + float w4 = 1.0; + float w1_3_5_7 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w0_2_6_8 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + vec3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + +// Resizable one-pass blurs: +vec3 tex2Dblur3x3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + float denom_inv = 0.5/(sigma*sigma); + float w0 = 1.0; + float w1 = exp(-1.0 * denom_inv); + float w2 = exp(-4.0 * denom_inv); + float w3 = exp(-9.0 * denom_inv); + float w4 = exp(-16.0 * denom_inv); + float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + float w12 = w1 + w2; + float w34 = w3 + w4; + float w12_ratio = w2/w12; + float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + vec3 sum = vec3(0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +vec3 tex2Dblur9x9(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float w4off = exp(-16.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2R_texel_offset = vec2(3.0, 0.0) + vec2(texel3to4ratio, 0.0); + vec2 sample3d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + vec2 sample4d_texel_offset = vec2(3.0, 1.0) + vec2(texel3to4ratio, texel1to2ratio); + vec2 sample5d_texel_offset = vec2(1.0, 3.0) + vec2(texel1to2ratio, texel3to4ratio); + vec2 sample6d_texel_offset = vec2(3.0, 3.0) + vec2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2R1 = w3off; + float w2R2 = w4off; + float w3d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w3d2_3d3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w3d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d1_5d1 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d2_5d3 = exp(-LENGTH_SQ(vec2(4.0, 1.0)) * denom_inv); + float w4d3_5d2 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4_5d4 = exp(-LENGTH_SQ(vec2(4.0, 2.0)) * denom_inv); + float w6d1 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + float w6d2_6d3 = exp(-LENGTH_SQ(vec2(4.0, 3.0)) * denom_inv); + float w6d4 = exp(-LENGTH_SQ(vec2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2R1 + w2R2; + float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + float w5 = w4; + float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + vec3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + vec3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + vec3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + vec3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + vec3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + vec3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + vec3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + vec3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + vec3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur7x7(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float w3off = exp(-9.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample1d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + vec2 sample2d_texel_offset = vec2(2.0, 0.0) + vec2(texel2to3ratio, texel0to1ratio); + vec2 sample3d_texel_offset = vec2(0.0, 2.0) + vec2(texel0to1ratio, texel2to3ratio); + vec2 sample4d_texel_offset = vec2(2.0, 2.0) + vec2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1abcd = 1.0; + float w1bd2_1cd3 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); + float w2bd1_3cd1 = exp(-LENGTH_SQ(vec2(2.0, 0.0)) * denom_inv); + float w2bd2_3cd2 = exp(-LENGTH_SQ(vec2(3.0, 0.0)) * denom_inv); + float w1d4 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d3_3d2 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4_3d4 = exp(-LENGTH_SQ(vec2(3.0, 1.0)) * denom_inv); + float w4d1 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + float w4d2_4d3 = exp(-LENGTH_SQ(vec2(3.0, 2.0)) * denom_inv); + float w4d4 = exp(-LENGTH_SQ(vec2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + vec3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + vec3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + vec3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + vec3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + vec3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + vec3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + vec3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + vec3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = vec3(0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur5x5(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w1off = exp(-1.0 * denom_inv); + float w2off = exp(-4.0 * denom_inv); + float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + vec2 sample1R_texel_offset = vec2(1.0, 0.0) + vec2(texel1to2ratio, 0.0); + vec2 sample2d_texel_offset = vec2(1.0, 1.0) + vec2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + float w1R1 = w1off; + float w1R2 = w2off; + float w2d1 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); + float w2d2_3 = exp(-LENGTH_SQ(vec2(2.0, 1.0)) * denom_inv); + float w2d4 = exp(-LENGTH_SQ(vec2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + float w0 = 1.0; + float w1 = w1R1 + w1R2; + float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + vec3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + vec3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + vec3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + vec3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + vec3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + vec3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + vec3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + vec3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +vec3 tex2Dblur3x3(sampler2D tex, vec2 tex_uv, + vec2 dxdy, float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + float denom_inv = 0.5/(sigma*sigma); + float w0off = 1.0; + float w1off = exp(-1.0 * denom_inv); + float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + vec2 sample0d_texel_offset = vec2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + vec2 mirror_x = vec2(-1.0, 1.0); + vec2 mirror_y = vec2(1.0, -1.0); + vec2 mirror_xy = vec2(-1.0, -1.0); + vec2 dxdy_mirror_x = dxdy * mirror_x; + vec2 dxdy_mirror_y = dxdy * mirror_y; + vec2 dxdy_mirror_xy = dxdy * mirror_xy; + vec3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + vec3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + vec3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + vec3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + +vec3 tex2Dblur9fast(sampler2D tex, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur17fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur17fast(texture, tex_uv, dxdy, blur17_std_dev); +} + +vec3 tex2Dblur25fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur25fast(texture, tex_uv, dxdy, blur25_std_dev); +} + +vec3 tex2Dblur43fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur43fast(texture, tex_uv, dxdy, blur43_std_dev); +} +vec3 tex2Dblur31fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur31fast(texture, tex_uv, dxdy, blur31_std_dev); +} + +vec3 tex2Dblur3fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3fast(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur3x3(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3x3(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5fast(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur5resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5resize(texture, tex_uv, dxdy, blur5_std_dev); +} +vec3 tex2Dblur3resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur3resize(texture, tex_uv, dxdy, blur3_std_dev); +} + +vec3 tex2Dblur5x5(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur5x5(texture, tex_uv, dxdy, blur5_std_dev); +} + +vec3 tex2Dblur7resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7resize(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7fast(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur7x7(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur7x7(texture, tex_uv, dxdy, blur7_std_dev); +} + +vec3 tex2Dblur9resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9resize(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur9x9(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur9x9(texture, tex_uv, dxdy, blur9_std_dev); +} + +vec3 tex2Dblur11resize(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11resize(texture, tex_uv, dxdy, blur11_std_dev); +} + +vec3 tex2Dblur11fast(sampler2D texture, vec2 tex_uv, + vec2 dxdy) +{ + return tex2Dblur11fast(texture, tex_uv, dxdy, blur11_std_dev); +} + +#endif // BLUR_FUNCTIONS_H + +#define InputSize sourceSize[0].xy +#define TextureSize sourceSize[0].xy +#define OutputSize targetSize.xy + +void main() { + gl_Position = position; + vTexCoord = texCoord; + // Get the uv sample distance between output pixels. Blurs are not generic + // Gaussian resizers, and correct blurs require: + // 1.) OutputSize == InputSize * 2^m, where m is an integer <= 0. + // 2.) mipmap_inputN = "true" for this pass in the preset if m != 0 + // 3.) filter_linearN = "true" except for 1x scale nearest neighbor blurs + // Gaussian resizers would upsize using the distance between input texels + // (not output pixels), but we avoid this and consistently blur at the + // destination size. Otherwise, combining statically calculated weights + // with bilinear sample exploitation would result in terrible artifacts. + vec2 dxdy_scale = InputSize/OutputSize; + vec2 dxdy = dxdy_scale/TextureSize; + // This blur is vertical-only, so zero out the horizontal offset: + blur_dxdy = vec2(0.0, dxdy.y); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/brightpass.fs b/shaders/CRT-Royale.shader/brightpass.fs new file mode 100644 index 000000000..29f27db8d --- /dev/null +++ b/shaders/CRT-Royale.shader/brightpass.fs @@ -0,0 +1,14481 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +in Vertex { + vec2 vTexCoord; + vec2 scanline_tex_uv; + vec2 blur3x3_tex_uv; + float bloom_sigma_runtime; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define MASKED_SCANLINEStexture source[0] +#define MASKED_SCANLINEStexture_size sourceSize[0].xy +#define MASKED_SCANLINESvideo_size sourceSize[0].xy +#define BLOOM_APPROXtexture source[5] +#define BLOOM_APPROXtexture_size sourceSize[5].xy +#define BLOOM_APPROXvideo_size sourceSize[5].xy + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +/////////////////////////////// END VERTEX-INCLUDES ///////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); +const float bloom_diff_thresh_ = 1.0/256.0; + +////////////////////////////// FRAGMENT INCLUDES ////////////////////////////// + +//#include "bloom-functions.h" + +//////////////////////////// BEGIN BLOOM-FUNCTIONS /////////////////////////// + +#ifndef BLOOM_FUNCTIONS_H +#define BLOOM_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These utility functions and constants help several passes determine the +// size and center texel weight of the phosphor bloom in a uniform manner. + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// We need to calculate the correct blur sigma using some .cgp constants: +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// output_size < video_size. +// 4.) output_size == video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (video_size/output_size)/texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(video_size/output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, output_size.x), [0, output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, output_size.x), [0, output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(video_size/output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +/////////////////////////////// BLOOM CONSTANTS ////////////////////////////// + +// Compute constants with manual inlines of the functions below: +static const float bloom_diff_thresh = 1.0/256.0; + + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +inline float get_absolute_scale_blur_sigma(const float thresh) +{ + // Requires: 1.) min_expected_triads must be a global float. The number + // of horizontal phosphor triads in the final image must be + // >= min_allowed_viewport_triads.x for realistic results. + // 2.) bloom_approx_scale_x must be a global float equal to the + // absolute horizontal scale of BLOOM_APPROX. + // 3.) bloom_approx_scale_x/min_allowed_viewport_triads.x + // should be <= 1.1658025090 to keep the final result < + // 0.62666015625 (the largest sigma ensuring the largest + // unused texel weight stays < 1.0/256.0 for a 3x3 blur). + // 4.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum Gaussian sigma that will blur the pass + // output as much as it would have taken to blur away + // bloom_approx_scale_x horizontal phosphor triads. + // Description: + // BLOOM_APPROX should look like a downscaled phosphor blur. Ideally, we'd + // use the same blur sigma as the actual phosphor bloom and scale it down + // to the current resolution with (bloom_approx_scale_x/viewport_size_x), but + // we don't know the viewport size in this pass. Instead, we'll blur as + // much as it would take to blur away min_allowed_viewport_triads.x. This + // will blur "more than necessary" if the user actually uses more triads, + // but that's not terrible either, because blurring a constant fraction of + // the viewport may better resemble a true optical bloom anyway (since the + // viewport will generally be about the same fraction of each player's + // field of view, regardless of screen size and resolution). + // Assume an extremely large viewport size for asymptotic results. + return bloom_approx_scale_x/max_viewport_size_x * + get_min_sigma_to_blur_triad( + max_viewport_size_x/min_allowed_viewport_triads.x, thresh); +} + +inline float get_center_weight(const float sigma) +{ + // Given a Gaussian blur sigma, get the blur weight for the center texel. + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return get_fast_gaussian_weight_sum_inv(sigma); + #else + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + // Note: If the implementation uses a smaller blur than the max allowed, + // the worst case scenario is that the center weight will be overestimated, + // so we'll put a bit more energy into the brightpass...no huge deal. + // Then again, if the implementation uses a larger blur than the max + // "allowed" because of dynamic branching, the center weight could be + // underestimated, which is more of a problem...consider always using + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // 43x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + + w11 + w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + // 31x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + + w8 + w9 + w10 + w11 + w12 + w13 + w14 + w15)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + // 25x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + // 17x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + #else + // 9x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + const float center_weight = weight_sum_inv * weight_sum_inv; + return center_weight; + #endif +} + +inline float3 tex2DblurNfast(const sampler2D texture, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // If sigma is static, we can safely branch and use the smallest blur + // that's big enough. Ignore #define hints, because we'll only use a + // large blur if we actually need it, and the branches cost nothing. + #ifndef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #else + // It's still worth branching if the profile supports dynamic branches: + // It's much faster than using a hugely excessive blur, but each branch + // eats ~1% FPS. + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #endif + #endif + // Failed optimization notes: + // I originally created a same-size mipmapped 5-tap separable blur10 that + // could handle any sigma by reaching into lower mip levels. It was + // as fast as blur25fast for runtime sigmas and a tad faster than + // blur31fast for static sigmas, but mipmapping two viewport-size passes + // ate 10% of FPS across all codepaths, so it wasn't worth it. + #ifdef PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + if(sigma <= blur9_std_dev) + { + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur17_std_dev) + { + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur25_std_dev) + { + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur31_std_dev) + { + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + } + else + { + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + } + #else + // If we can't afford to branch, we can only guess at what blur + // size we need. Therefore, use the largest blur allowed. + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + #else + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + #endif // PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE +} + +inline float get_bloom_approx_sigma(const float output_size_x_runtime, + const float estimated_viewport_size_x) +{ + // Requires: 1.) output_size_x_runtime == BLOOM_APPROX.output_size.x. + // This is included for dynamic codepaths just in case the + // following two globals are incorrect: + // 2.) bloom_approx_size_x_for_skip should == the same + // if PHOSPHOR_BLOOM_FAKE is #defined + // 3.) bloom_approx_size_x should == the same otherwise + // Returns: For gaussian4x4, return a dynamic small bloom sigma that's + // as close to optimal as possible given available information. + // For blur3x3, return the a static small bloom sigma that + // works well for typical cases. Otherwise, we're using simple + // bilinear filtering, so use static calculations. + // Assume the default static value. This is a compromise that ensures + // typical triads are blurred, even if unusually large ones aren't. + static const float mask_num_triads_static = + max(min_allowed_viewport_triads.x, mask_num_triads_desired_static); + const float mask_num_triads_from_size = + estimated_viewport_size_x/mask_triad_size_desired; + const float mask_num_triads_runtime = max(min_allowed_viewport_triads.x, + lerp(mask_num_triads_from_size, mask_num_triads_desired, + mask_specify_num_triads)); + // Assume an extremely large viewport size for asymptotic results: + static const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + if(bloom_approx_filter > 1.5) // 4x4 true Gaussian resize + { + // Use the runtime num triads and output size: + const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_runtime; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_runtime/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // account for the Gaussian scanline sigma from the last pass too. + // The bloom will be too wide horizontally but tall enough vertically. + return length(float2(bloom_approx_sigma, beam_max_sigma)); + } + else // 3x3 blur resize (the bilinear resize doesn't need a sigma) + { + // We're either using blur3x3 or bilinear filtering. The biggest + // reason to choose blur3x3 is to avoid dynamic weights, so use a + // static calculation. + #ifdef PHOSPHOR_BLOOM_FAKE + static const float output_size_x_static = + bloom_approx_size_x_for_fake; + #else + static const float output_size_x_static = bloom_approx_size_x; + #endif + static const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_static; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_static/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // try accounting for the Gaussian scanline sigma from the last pass + // too; use the static default value: + return length(float2(bloom_approx_sigma, beam_max_sigma_static)); + } +} + +inline float get_final_bloom_sigma(const float bloom_sigma_runtime) +{ + // Requires: 1.) bloom_sigma_runtime is a precalculated sigma that's + // optimal for the [known] triad size. + // 2.) Call this from a fragment shader (not a vertex shader), + // or blurring with static sigmas won't be constant-folded. + // Returns: Return the optimistic static sigma if the triad size is + // known at compile time. Otherwise return the optimal runtime + // sigma (10% slower) or an implementation-specific compromise + // between an optimistic or pessimistic static sigma. + // Notes: Call this from the fragment shader, NOT the vertex shader, + // so static sigmas can be constant-folded! + const float bloom_sigma_optimistic = get_min_sigma_to_blur_triad( + mask_triad_size_desired_static, bloom_diff_thresh); + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return bloom_sigma_runtime; + #else + // Overblurring looks as bad as underblurring, so assume average-size + // triads, not worst-case huge triads: + return bloom_sigma_optimistic; + #endif +} + + +#endif // BLOOM_FUNCTIONS_H + +//////////////////////////// END BLOOM-FUNCTIONS /////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// output_size < video_size. +// 4.) output_size == video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (video_size/output_size)/texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(video_size/output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, output_size.x), [0, output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, output_size.x), [0, output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(video_size/output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +void main() { + // Sample the masked scanlines: + const float3 intensity_dim = + tex2D_linearize(MASKED_SCANLINEStexture, scanline_tex_uv).rgb; + // Get the full intensity, including auto-undimming, and mask compensation: + const float auto_dim_factor = levels_autodim_temp; + const float undim_factor = 1.0/auto_dim_factor; + const float mask_amplify = get_mask_amplify(); + const float3 intensity = intensity_dim * undim_factor * mask_amplify * + levels_contrast; + + // Sample BLOOM_APPROX to estimate what a straight blur of masked scanlines + // would look like, so we can estimate how much energy we'll receive from + // blooming neighbors: + const float3 phosphor_blur_approx = levels_contrast * tex2D_linearize( + BLOOM_APPROXtexture, blur3x3_tex_uv).rgb; + + // Compute the blur weight for the center texel and the maximum energy we + // expect to receive from neighbors: + const float bloom_sigma = get_final_bloom_sigma(bloom_sigma_runtime); + const float center_weight = get_center_weight(bloom_sigma); + const float3 max_area_contribution_approx = + max(float3(0.0, 0.0, 0.0), phosphor_blur_approx - center_weight * intensity); + // Assume neighbors will blur 100% of their intensity (blur_ratio = 1.0), + // because it actually gets better results (on top of being very simple), + // but adjust all intensities for the user's desired underestimate factor: + const float3 area_contrib_underestimate = + bloom_underestimate_levels * max_area_contribution_approx; + const float3 intensity_underestimate = + bloom_underestimate_levels * intensity; + // Calculate the blur_ratio, the ratio of intensity we want to blur: + #ifdef BRIGHTPASS_AREA_BASED + // This area-based version changes blur_ratio more smoothly and blurs + // more, clipping less but offering less phosphor differentiation: + const float3 phosphor_blur_underestimate = bloom_underestimate_levels * + phosphor_blur_approx; + const float3 soft_intensity = max(intensity_underestimate, + phosphor_blur_underestimate * mask_amplify); + const float3 blur_ratio_temp = + ((float3(1.0, 1.0, 1.0) - area_contrib_underestimate) / + soft_intensity - float3(1.0, 1.0, 1.0)) / (center_weight - 1.0); + #else + const float3 blur_ratio_temp = + ((float3(1.0, 1.0, 1.0) - area_contrib_underestimate) / + intensity_underestimate - float3(1.0, 1.0, 1.0)) / (center_weight - 1.0); + #endif + const float3 blur_ratio = clamp(blur_ratio_temp, 0.0, 1.0); + // Calculate the brightpass based on the auto-dimmed, unamplified, masked + // scanlines, encode if necessary, and return! + const float3 brightpass = intensity_dim * + lerp(blur_ratio, float3(1.0, 1.0, 1.0), bloom_excess); + FragColor = encode_output(float4(brightpass, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/brightpass.vs b/shaders/CRT-Royale.shader/brightpass.vs new file mode 100644 index 000000000..2d02d72a8 --- /dev/null +++ b/shaders/CRT-Royale.shader/brightpass.vs @@ -0,0 +1,6551 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 scanline_tex_uv; + vec2 blur3x3_tex_uv; + float bloom_sigma_runtime; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define MASKED_SCANLINEStexture source[0] +#define MASKED_SCANLINEStexture_size sourceSize[0].xy +#define MASKED_SCANLINESvideo_size sourceSize[0].xy +#define BLOOM_APPROXtexture source[3] +#define BLOOM_APPROXtexture_size sourceSize[3].xy +#define BLOOM_APPROXvideo_size sourceSize[3].xy + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +/////////////////////////////// END VERTEX-INCLUDES ///////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); +const float bloom_diff_thresh_ = 1.0/256.0; + +// copied from bloom-functions.h +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +void main() { + gl_Position = position; + vTexCoord = texCoord * 1.00001; + float2 tex_uv = vTexCoord.xy; + // Our various input textures use different coords: + float2 video_uv = tex_uv * texture_size/video_size; + //video_uv = video_uv; + scanline_tex_uv = video_uv * MASKED_SCANLINESvideo_size / + MASKED_SCANLINEStexture_size; + blur3x3_tex_uv = video_uv;// * BLOOM_APPROXvideo_size / BLOOM_APPROXtexture_size; + + // Calculate a runtime bloom_sigma in case it's needed: + const float mask_tile_size_x = get_resized_mask_tile_size( + output_size, output_size * mask_resize_viewport_scale, false).x; + bloom_sigma_runtime = get_min_sigma_to_blur_triad( + mask_tile_size_x / mask_triads_per_tile, bloom_diff_thresh_); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/first-pass-linearize-crt-gamma-bob-fields.fs b/shaders/CRT-Royale.shader/first-pass-linearize-crt-gamma-bob-fields.fs new file mode 100644 index 000000000..c89e46716 --- /dev/null +++ b/shaders/CRT-Royale.shader/first-pass-linearize-crt-gamma-bob-fields.fs @@ -0,0 +1,4748 @@ +#version 150 + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; +uniform int phase; + +in Vertex { + vec2 vTexCoord; + vec2 uv_step; + float interlaced; +}; + +out vec4 FragColor; + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#ifdef GL_ES +#ifdef GL_FRAGMENT_PRECISION_HIGH +precision highp float; +#else +precision mediump float; +#endif +#define COMPAT_PRECISION mediump +#else +#define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 +#define COMPAT_VARYING in +#define COMPAT_TEXTURE texture +#else +#define COMPAT_VARYING varying +#define FragColor gl_FragColor +#define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// PASS SETTINGS: +// gamma-management.h needs to know what kind of pipeline we're using and +// what pass this is in that pipeline. This will become obsolete if/when we +// can #define things like this in the .cgp preset file. +#define FIRST_PASS +#define SIMULATE_CRT_ON_LCD + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +void main() { + const float2 tex_uv = vTexCoord.xy; + // Linearize the input based on CRT gamma and bob interlaced fields. + // Bobbing ensures we can immediately blur without getting artifacts. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + if(bool(interlace_detect)) + { + // Sample the current line and an average of the previous/next line; + // tex2D_linearize will decode CRT gamma. Don't bother branching: + const float2 v_step = float2(0.0, uv_step.y); + const float3 curr_line = tex2D_linearize( + input_texture, tex_uv).rgb; + const float3 last_line = tex2D_linearize( + input_texture, tex_uv - v_step).rgb; + const float3 next_line = tex2D_linearize( + input_texture, tex_uv + v_step).rgb; + const float3 interpolated_line = 0.5 * (last_line + next_line); + // If we're interlacing, determine which field curr_line is in: + const float modulus = interlaced + 1.0; + const float field_offset = + fmod(frame_count + interlace_bff, modulus); + const float curr_line_texel = tex_uv.y * texture_size.y; + // Use under_half to fix a rounding bug around exact texel locations. + const float line_num_last = floor(curr_line_texel - under_half); + const float wrong_field = fmod(line_num_last + field_offset, modulus); + // Select the correct color, and output the result: + const float3 color = lerp(curr_line, interpolated_line, wrong_field); + FragColor = encode_output(float4(color, 1.0)); + } + else + { + FragColor = encode_output(tex2D_linearize(input_texture, tex_uv)); + } +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/first-pass-linearize-crt-gamma-bob-fields.vs b/shaders/CRT-Royale.shader/first-pass-linearize-crt-gamma-bob-fields.vs new file mode 100644 index 000000000..12b93534e --- /dev/null +++ b/shaders/CRT-Royale.shader/first-pass-linearize-crt-gamma-bob-fields.vs @@ -0,0 +1,4704 @@ +#version 150 + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 uv_step; + float interlaced; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#ifdef GL_ES +#ifdef GL_FRAGMENT_PRECISION_HIGH +precision highp float; +#else +precision mediump float; +#endif +#define COMPAT_PRECISION mediump +#else +#define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 +#define COMPAT_VARYING in +#define COMPAT_TEXTURE texture +#else +#define COMPAT_VARYING varying +#define FragColor gl_FragColor +#define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// PASS SETTINGS: +// gamma-management.h needs to know what kind of pipeline we're using and +// what pass this is in that pipeline. This will become obsolete if/when we +// can #define things like this in the .cgp preset file. +#define FIRST_PASS +#define SIMULATE_CRT_ON_LCD + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +void main() { + gl_Position = position; + vTexCoord = texCoord; + uv_step = float2(1.0)/texture_size; + + // Detect interlacing: 1.0 = true, 0.0 = false. + const float2 _video_size = video_size; + interlaced = float(is_interlaced(_video_size.y)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/geometry-aa-last-pass.fs b/shaders/CRT-Royale.shader/geometry-aa-last-pass.fs new file mode 100644 index 000000000..87d1b7213 --- /dev/null +++ b/shaders/CRT-Royale.shader/geometry-aa-last-pass.fs @@ -0,0 +1,5279 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; +uniform int phase; + +in Vertex { + vec2 vTexCoord; + vec2 tex_uv; + vec4 video_and_texture_size_inv; + vec2 output_size_inv; + vec3 eye_pos_local; + vec4 geom_aspect_and_overscan; + vec3 global_to_local_row0; + vec3 global_to_local_row1; + vec3 global_to_local_row2; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 1.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(x,y) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +#define LAST_PASS +#define SIMULATE_CRT_ON_LCD + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +#ifndef RUNTIME_GEOMETRY_TILT + // Create a local-to-global rotation matrix for the CRT's coordinate frame + // and its global-to-local inverse. See the vertex shader for details. + // It's faster to compute these statically if possible. + static const float2 sin_tilt = sin(geom_tilt_angle_static); + static const float2 cos_tilt = cos(geom_tilt_angle_static); + static const float3x3 geom_local_to_global_static = float3x3( + cos_tilt.x, sin_tilt.y*sin_tilt.x, cos_tilt.y*sin_tilt.x, + 0.0, cos_tilt.y, -sin_tilt.y, + -sin_tilt.x, sin_tilt.y*cos_tilt.x, cos_tilt.y*cos_tilt.x); + static const float3x3 geom_global_to_local_static = float3x3( + cos_tilt.x, 0.0, -sin_tilt.x, + sin_tilt.y*sin_tilt.x, cos_tilt.y, sin_tilt.y*cos_tilt.x, + cos_tilt.y*sin_tilt.x, -sin_tilt.y, cos_tilt.y*cos_tilt.x); +#endif + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "tex2Dantialias.h" + +///////////////////////// BEGIN TEX2DANTIALIAS ///////////////////////// + +#ifndef TEX2DANTIALIAS_H +#define TEX2DANTIALIAS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides antialiased and subpixel-aware tex2D lookups. +// Requires: All functions share these requirements: +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) pixel_to_tex_uv must be a 2x2 matrix that transforms pixe- +// space offsets to texture uv offsets. You can get this with: +// const float2 duv_dx = ddx(tex_uv); +// const float2 duv_dy = ddy(tex_uv); +// const float2x2 pixel_to_tex_uv = float2x2( +// duv_dx.x, duv_dy.x, +// duv_dx.y, duv_dy.y); +// This is left to the user in case the current Cg profile +// doesn't support ddx()/ddy(). Ideally, the user could find +// calculate a distorted tangent-space mapping analytically. +// If not, a simple flat mapping can be obtained with: +// const float2 xy_to_uv_scale = output_size * +// video_size/texture_size; +// const float2x2 pixel_to_tex_uv = float2x2( +// xy_to_uv_scale.x, 0.0, +// 0.0, xy_to_uv_scale.y); +// Optional: To set basic AA settings, #define ANTIALIAS_OVERRIDE_BASICS and: +// 1.) Set an antialiasing level: +// static const float aa_level = {0 (none), +// 1 (sample subpixels), 4, 5, 6, 7, 8, 12, 16, 20, 24} +// 2.) Set a filter type: +// static const float aa_filter = { +// 0 (Box, Separable), 1 (Box, Cylindrical), +// 2 (Tent, Separable), 3 (Tent, Cylindrical) +// 4 (Gaussian, Separable), 5 (Gaussian, Cylindrical) +// 6 (Cubic, Separable), 7 (Cubic, Cylindrical) +// 8 (Lanczos Sinc, Separable), +// 9 (Lanczos Jinc, Cylindrical)} +// If the input is unknown, a separable box filter is used. +// Note: Lanczos Jinc is terrible for sparse sampling, and +// using aa_axis_importance (see below) defeats the purpose. +// 3.) Mirror the sample pattern on odd frames? +// static const bool aa_temporal = {true, false] +// This helps rotational invariance but can look "fluttery." +// The user may #define ANTIALIAS_OVERRIDE_PARAMETERS to override +// (all of) the following default parameters with static or uniform +// constants (or an accessor function for subpixel offsets): +// 1.) Cubic parameters: +// static const float aa_cubic_c = 0.5; +// See http://www.imagemagick.org/Usage/filter/#mitchell +// 2.) Gaussian parameters: +// static const float aa_gauss_sigma = +// 0.5/aa_pixel_diameter; +// 3.) Set subpixel offsets. This requires an accessor function +// for compatibility with scalar runtime shader Return +// a float2 pixel offset in [-0.5, 0.5] for the red subpixel: +// float2 get_aa_subpixel_r_offset() +// The user may also #define ANTIALIAS_OVERRIDE_STATIC_CONSTANTS to +// override (all of) the following default static values. However, +// the file's structure requires them to be declared static const: +// 1.) static const float aa_lanczos_lobes = 3.0; +// 2.) static const float aa_gauss_support = 1.0/aa_pixel_diameter; +// Note the default tent/Gaussian support radii may appear +// arbitrary, but extensive testing found them nearly optimal +// for tough cases like strong distortion at low AA levels. +// (The Gaussian default is only best for practical gauss_sigma +// values; much larger gauss_sigmas ironically prefer slightly +// smaller support given sparse sampling, and vice versa.) +// 3.) static const float aa_tent_support = 1.0 / aa_pixel_diameter; +// 4.) static const float2 aa_xy_axis_importance: +// The sparse N-queens sampling grid interacts poorly with +// negative-lobed 2D filters. However, if aliasing is much +// stronger in one direction (e.g. horizontally with a phosphor +// mask), it can be useful to downplay sample offsets along the +// other axis. The support radius in each direction scales with +// aa_xy_axis_importance down to a minimum of 0.5 (box support), +// after which point only the offsets used for calculating +// weights continue to scale downward. This works as follows: +// If aa_xy_axis_importance = float2(1.0, 1.0/support_radius), +// the vertical support radius will drop to 1.0, and we'll just +// filter vertical offsets with the first filter lobe, while +// horizontal offsets go through the full multi-lobe filter. +// If aa_xy_axis_importance = float2(1.0, 0.0), the vertical +// support radius will drop to box support, and the vertical +// offsets will be ignored entirely (essentially giving us a +// box filter vertically). The former is potentially smoother +// (but less predictable) and the default behavior of Lanczos +// jinc, whereas the latter is sharper and the default behavior +// of cubics and Lanczos sinc. +// 5.) static const float aa_pixel_diameter: You can expand the +// pixel diameter to e.g. sqrt(2.0), which may be a better +// support range for cylindrical filters (they don't +// currently discard out-of-circle samples though). +// Finally, there are two miscellaneous options: +// 1.) If you want to antialias a manually tiled texture, you can +// #define ANTIALIAS_DISABLE_ANISOTROPIC to use tex2Dlod() to +// fix incompatibilities with anisotropic filtering. This is +// slower, and the Cg profile must support tex2Dlod(). +// 2.) If aa_cubic_c is a runtime uniform, you can #define +// RUNTIME_ANTIALIAS_WEIGHTS to evaluate cubic weights once per +// fragment instead of at the usage site (which is used by +// default, because it enables static evaluation). +// Description: +// Each antialiased lookup follows these steps: +// 1.) Define a sample pattern of pixel offsets in the range of [-0.5, 0.5] +// pixels, spanning the diameter of a rectangular box filter. +// 2.) Scale these offsets by the support diameter of the user's chosen filter. +// 3.) Using these pixel offsets from the pixel center, compute the offsets to +// predefined subpixel locations. +// 4.) Compute filter weights based on subpixel offsets. +// Much of that can often be done at compile-time. At runtime: +// 1.) Project pixel-space offsets into uv-space with a matrix multiplication +// to get the uv offsets for each sample. Rectangular pixels have a +// diameter of 1.0. Circular pixels are not currently supported, but they +// might be better with a diameter of sqrt(2.0) to ensure there are no gaps +// between them. +// 2.) Load, weight, and sum samples. +// We use a sparse bilinear sampling grid, so there are two major implications: +// 1.) We can directly project the pixel-space support box into uv-space even +// if we're upsizing. This wouldn't be the case for nearest neighbor, +// where we'd have to expand the uv-space diameter to at least the support +// size to ensure sufficient filter support. In our case, this allows us +// to treat upsizing the same as downsizing and use static weighting. :) +// 2.) For decent results, negative-lobed filters must be computed based on +// separable weights, not radial distances, because the sparse sampling +// makes no guarantees about radial distributions. Even then, it's much +// better to set aa_xy_axis_importance to e.g. float2(1.0, 0.0) to use e.g. +// Lanczos2 horizontally and a box filter vertically. This is mainly due +// to the sparse N-queens sampling and a statistically enormous positive or +// negative covariance between horizontal and vertical weights. +// +// Design Decision Comments: +// "aa_temporal" mirrors the sample pattern on odd frames along the axis that +// keeps subpixel weights constant. This helps with rotational invariance, but +// it can cause distracting fluctuations, and horizontal and vertical edges +// will look the same. Using a different pattern on a shifted grid would +// exploit temporal AA better, but it would require a dynamic branch or a lot +// of conditional moves, so it's prohibitively slow for the minor benefit. + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +#ifndef ANTIALIAS_OVERRIDE_BASICS + // The following settings must be static constants: + static const float aa_level = 12.0; + static const float aa_filter = 0.0; + static const bool aa_temporal = false; +#endif + +#ifndef ANTIALIAS_OVERRIDE_STATIC_CONSTANTS + // Users may override these parameters, but the file structure requires + // them to be static constants; see the descriptions above. + static const float aa_pixel_diameter = 1.0; + static const float aa_lanczos_lobes = 3.0; + static const float aa_gauss_support = 1.0 / aa_pixel_diameter; + static const float aa_tent_support = 1.0 / aa_pixel_diameter; + + // If we're using a negative-lobed filter, default to using it horizontally + // only, and use only the first lobe vertically or a box filter, over a + // correspondingly smaller range. This compensates for the sparse sampling + // grid's typically large positive/negative x/y covariance. + static const float2 aa_xy_axis_importance = + aa_filter < 5.5 ? float2(1.0) : // Box, tent, Gaussian + aa_filter < 8.5 ? float2(1.0, 0.0) : // Cubic and Lanczos sinc + aa_filter < 9.5 ? float2(1.0, 1.0/aa_lanczos_lobes) : // Lanczos jinc + float2(1.0); // Default to box +#endif + +#ifndef ANTIALIAS_OVERRIDE_PARAMETERS + // Users may override these values with their own uniform or static consts. + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c = 0.5; + static const float aa_gauss_sigma = 0.5 / aa_pixel_diameter; + // Users may override the subpixel offset accessor function with their own. + // A function is used for compatibility with scalar runtime shader + inline float2 get_aa_subpixel_r_offset() + { + return float2(0.0, 0.0); + } +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../../../../include/gamma-management.h" + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +static const float aa_box_support = 0.5; +static const float aa_cubic_support = 2.0; + + +//////////////////////////// GLOBAL NON-CONSTANTS //////////////////////////// + +// We'll want to define these only once per fragment at most. +#ifdef RUNTIME_ANTIALIAS_WEIGHTS + float aa_cubic_b; + float cubic_branch1_x3_coeff; + float cubic_branch1_x2_coeff; + float cubic_branch1_x0_coeff; + float cubic_branch2_x3_coeff; + float cubic_branch2_x2_coeff; + float cubic_branch2_x1_coeff; + float cubic_branch2_x0_coeff; +#endif + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +void assign_aa_cubic_constants() +{ + // Compute cubic coefficients on demand at runtime, and save them to global + // uniforms. The B parameter is computed from C, because "Keys cubics" + // with B = 1 - 2C are considered the highest quality. + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + if(aa_filter > 5.5 && aa_filter < 7.5) + { + aa_cubic_b = 1.0 - 2.0*aa_cubic_c; + cubic_branch1_x3_coeff = 12.0 - 9.0*aa_cubic_b - 6.0*aa_cubic_c; + cubic_branch1_x2_coeff = -18.0 + 12.0*aa_cubic_b + 6.0*aa_cubic_c; + cubic_branch1_x0_coeff = 6.0 - 2.0 * aa_cubic_b; + cubic_branch2_x3_coeff = -aa_cubic_b - 6.0 * aa_cubic_c; + cubic_branch2_x2_coeff = 6.0*aa_cubic_b + 30.0*aa_cubic_c; + cubic_branch2_x1_coeff = -12.0*aa_cubic_b - 48.0*aa_cubic_c; + cubic_branch2_x0_coeff = 8.0*aa_cubic_b + 24.0*aa_cubic_c; + } + #endif +} + +inline float4 get_subpixel_support_diam_and_final_axis_importance() +{ + // Statically select the base support radius: + static const float base_support_radius = + aa_filter < 1.5 ? aa_box_support : + aa_filter < 3.5 ? aa_tent_support : + aa_filter < 5.5 ? aa_gauss_support : + aa_filter < 7.5 ? aa_cubic_support : + aa_filter < 9.5 ? aa_lanczos_lobes : + aa_box_support; // Default to box + // Expand the filter support for subpixel filtering. + const float2 subpixel_support_radius_raw = + float2(base_support_radius) + abs(get_aa_subpixel_r_offset()); + if(aa_filter < 1.5) + { + // Ignore aa_xy_axis_importance for box filtering. + const float2 subpixel_support_diam = + 2.0 * subpixel_support_radius_raw; + const float2 final_axis_importance = float2(1.0); + return float4(subpixel_support_diam, final_axis_importance); + } + else + { + // Scale the support window by aa_xy_axis_importance, but don't narrow + // it further than box support. This allows decent vertical AA without + // messing up horizontal weights or using something silly like Lanczos4 + // horizontally with a huge vertical average over an 8-pixel radius. + const float2 subpixel_support_radius = max(float2(aa_box_support, aa_box_support), + subpixel_support_radius_raw * aa_xy_axis_importance); + // Adjust aa_xy_axis_importance to compensate for what's already done: + const float2 final_axis_importance = aa_xy_axis_importance * + subpixel_support_radius_raw/subpixel_support_radius; + const float2 subpixel_support_diam = 2.0 * subpixel_support_radius; + return float4(subpixel_support_diam, final_axis_importance); + } +} + + +/////////////////////////// FILTER WEIGHT FUNCTIONS ////////////////////////// + +inline float eval_box_filter(const float dist) +{ + return float(abs(dist) <= aa_box_support); +} + +inline float eval_separable_box_filter(const float2 offset) +{ + return float(all(bool2((abs(offset.x) <= aa_box_support), (abs(offset.y) <= aa_box_support)))); +} + +inline float eval_tent_filter(const float dist) +{ + return clamp((aa_tent_support - dist)/ + aa_tent_support, 0.0, 1.0); +} + +inline float eval_gaussian_filter(const float dist) +{ + return exp(-(dist*dist) / (2.0*aa_gauss_sigma*aa_gauss_sigma)); +} + +inline float eval_cubic_filter(const float dist) +{ + // Compute coefficients like assign_aa_cubic_constants(), but statically. + #ifndef RUNTIME_ANTIALIAS_WEIGHTS + // When runtime weights are used, these values are instead written to + // global uniforms at the beginning of each tex2Daa* call. + const float aa_cubic_b = 1.0 - 2.0*aa_cubic_c; + const float cubic_branch1_x3_coeff = 12.0 - 9.0*aa_cubic_b - 6.0*aa_cubic_c; + const float cubic_branch1_x2_coeff = -18.0 + 12.0*aa_cubic_b + 6.0*aa_cubic_c; + const float cubic_branch1_x0_coeff = 6.0 - 2.0 * aa_cubic_b; + const float cubic_branch2_x3_coeff = -aa_cubic_b - 6.0 * aa_cubic_c; + const float cubic_branch2_x2_coeff = 6.0*aa_cubic_b + 30.0*aa_cubic_c; + const float cubic_branch2_x1_coeff = -12.0*aa_cubic_b - 48.0*aa_cubic_c; + const float cubic_branch2_x0_coeff = 8.0*aa_cubic_b + 24.0*aa_cubic_c; + #endif + const float abs_dist = abs(dist); + // Compute the cubic based on the Horner's method formula in: + // http://www.cs.utexas.edu/users/fussell/courses/cs384g/lectures/mitchell/Mitchell.pdf + return (abs_dist < 1.0 ? + (cubic_branch1_x3_coeff*abs_dist + + cubic_branch1_x2_coeff)*abs_dist*abs_dist + + cubic_branch1_x0_coeff : + abs_dist < 2.0 ? + ((cubic_branch2_x3_coeff*abs_dist + + cubic_branch2_x2_coeff)*abs_dist + + cubic_branch2_x1_coeff)*abs_dist + cubic_branch2_x0_coeff : + 0.0)/6.0; +} + +inline float eval_separable_cubic_filter(const float2 offset) +{ + // This is faster than using a specific float2 version: + return eval_cubic_filter(offset.x) * + eval_cubic_filter(offset.y); +} + +inline float2 eval_sinc_filter(const float2 offset) +{ + // It's faster to let the caller handle the zero case, or at least it + // was when I used macros and the shader preset took a full minute to load. + const float2 pi_offset = pi * offset; + return sin(pi_offset)/pi_offset; +} + +inline float eval_separable_lanczos_sinc_filter(const float2 offset_unsafe) +{ + // Note: For sparse sampling, you really need to pick an axis to use + // Lanczos along (e.g. set aa_xy_axis_importance = float2(1.0, 0.0)). + const float2 offset = FIX_ZERO(offset_unsafe); + const float2 xy_weights = eval_sinc_filter(offset) * + eval_sinc_filter(offset/aa_lanczos_lobes); + return xy_weights.x * xy_weights.y; +} + +inline float eval_jinc_filter_unorm(const float x) +{ + // This is a Jinc approximation for x in [0, 45). We'll use x in range + // [0, 4*pi) or so. There are faster/closer approximations based on + // piecewise cubics from [0, 45) and asymptotic approximations beyond that, + // but this has a maximum absolute error < 1/512, and it's simpler/faster + // for shaders...not that it's all that useful for sparse sampling anyway. + const float point3845_x = 0.38448566093564*x; + const float exp_term = exp(-(point3845_x*point3845_x)); + const float point8154_plus_x = 0.815362332840791 + x; + const float cos_term = cos(point8154_plus_x); + return ( + 0.0264727330997042*min(x, 6.83134964622778) + + 0.680823557250528*exp_term + + -0.0597255978950933*min(7.41043194481873, x)*cos_term / + (point8154_plus_x + 0.0646074538634482*(x*x) + + cos(x)*max(exp_term, cos(x) + cos_term)) - + 0.180837503591406); +} + +inline float eval_jinc_filter(const float dist) +{ + return eval_jinc_filter_unorm(pi * dist); +} + +inline float eval_lanczos_jinc_filter(const float dist) +{ + return eval_jinc_filter(dist) * eval_jinc_filter(dist/aa_lanczos_lobes); +} + + +inline float3 eval_unorm_rgb_weights(const float2 offset, + const float2 final_axis_importance) +{ + // Requires: 1.) final_axis_impportance must be computed according to + // get_subpixel_support_diam_and_final_axis_importance(). + // 2.) aa_filter must be a global constant. + // 3.) offset must be an xy pixel offset in the range: + // ([-subpixel_support_diameter.x/2, + // subpixel_support_diameter.x/2], + // [-subpixel_support_diameter.y/2, + // subpixel_support_diameter.y/2]) + // Returns: Sample weights at R/G/B destination subpixels for the + // given xy pixel offset. + const float2 offset_g = offset * final_axis_importance; + const float2 aa_r_offset = get_aa_subpixel_r_offset(); + const float2 offset_r = offset_g - aa_r_offset * final_axis_importance; + const float2 offset_b = offset_g + aa_r_offset * final_axis_importance; + // Statically select a filter: + if(aa_filter < 0.5) + { + return float3(eval_separable_box_filter(offset_r), + eval_separable_box_filter(offset_g), + eval_separable_box_filter(offset_b)); + } + else if(aa_filter < 1.5) + { + return float3(eval_box_filter(length(offset_r)), + eval_box_filter(length(offset_g)), + eval_box_filter(length(offset_b))); + } + else if(aa_filter < 2.5) + { + return float3( + eval_tent_filter(offset_r.x) * eval_tent_filter(offset_r.y), + eval_tent_filter(offset_g.x) * eval_tent_filter(offset_g.y), + eval_tent_filter(offset_b.x) * eval_tent_filter(offset_b.y)); + } + else if(aa_filter < 3.5) + { + return float3(eval_tent_filter(length(offset_r)), + eval_tent_filter(length(offset_g)), + eval_tent_filter(length(offset_b))); + } + else if(aa_filter < 4.5) + { + return float3( + eval_gaussian_filter(offset_r.x) * eval_gaussian_filter(offset_r.y), + eval_gaussian_filter(offset_g.x) * eval_gaussian_filter(offset_g.y), + eval_gaussian_filter(offset_b.x) * eval_gaussian_filter(offset_b.y)); + } + else if(aa_filter < 5.5) + { + return float3(eval_gaussian_filter(length(offset_r)), + eval_gaussian_filter(length(offset_g)), + eval_gaussian_filter(length(offset_b))); + } + else if(aa_filter < 6.5) + { + return float3( + eval_cubic_filter(offset_r.x) * eval_cubic_filter(offset_r.y), + eval_cubic_filter(offset_g.x) * eval_cubic_filter(offset_g.y), + eval_cubic_filter(offset_b.x) * eval_cubic_filter(offset_b.y)); + } + else if(aa_filter < 7.5) + { + return float3(eval_cubic_filter(length(offset_r)), + eval_cubic_filter(length(offset_g)), + eval_cubic_filter(length(offset_b))); + } + else if(aa_filter < 8.5) + { + return float3(eval_separable_lanczos_sinc_filter(offset_r), + eval_separable_lanczos_sinc_filter(offset_g), + eval_separable_lanczos_sinc_filter(offset_b)); + } + else if(aa_filter < 9.5) + { + return float3(eval_lanczos_jinc_filter(length(offset_r)), + eval_lanczos_jinc_filter(length(offset_g)), + eval_lanczos_jinc_filter(length(offset_b))); + } + else + { + // Default to a box, because Lanczos Jinc is so bad. ;) + return float3(eval_separable_box_filter(offset_r), + eval_separable_box_filter(offset_g), + eval_separable_box_filter(offset_b)); + } +} + + +////////////////////////////// HELPER FUNCTIONS ////////////////////////////// + +inline float4 tex2Daa_tiled_linearize(const sampler2D samp, const float2 s) +{ + // If we're manually tiling a texture, anisotropic filtering can get + // confused. This is one workaround: + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + // TODO: Use tex2Dlod_linearize with a calculated mip level. + return tex2Dlod_linearize(samp, float4(s, 0.0, 0.0)); + #else + return tex2D_linearize(samp, s); + #endif +} + +inline float2 get_frame_sign(const float frame) +{ + if(aa_temporal) + { + // Mirror the sampling pattern for odd frames in a direction that + // lets us keep the same subpixel sample weights: + const float frame_odd = float(fmod(frame, 2.0) > 0.5); + const float2 aa_r_offset = get_aa_subpixel_r_offset(); + const float2 mirror = -float2(abs(aa_r_offset.x) < (FIX_ZERO(0.0)), abs(aa_r_offset.y) < (FIX_ZERO(0.0))); + return mirror; + } + else + { + return float2(1.0, 1.0); + } +} + + +///////////////////////// ANTIALIASED TEXTURE LOOKUPS //////////////////////// + +float3 tex2Daa_subpixel_weights_only(const sampler2D tex, + const float2 tex_uv, const float2x2 pixel_to_tex_uv) +{ + // This function is unlike the others: Just perform a single independent + // lookup for each subpixel. It may be very aliased. + const float2 aa_r_offset = get_aa_subpixel_r_offset(); + const float2 aa_r_offset_uv_offset = mul(pixel_to_tex_uv, aa_r_offset); + const float color_g = tex2D_linearize(tex, tex_uv).g; + const float color_r = tex2D_linearize(tex, tex_uv + aa_r_offset_uv_offset).r; + const float color_b = tex2D_linearize(tex, tex_uv - aa_r_offset_uv_offset).b; + return float3(color_r, color_g, color_b); +} + +// The tex2Daa* functions compile very slowly due to all the macros and +// compile-time math, so only include the ones we'll actually use! +float3 tex2Daa4x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use an RGMS4 pattern (4-queens): + // . . Q . : off =(-1.5, -1.5)/4 + (2.0, 0.0)/4 + // Q . . . : off =(-1.5, -1.5)/4 + (0.0, 1.0)/4 + // . . . Q : off =(-1.5, -1.5)/4 + (3.0, 2.0)/4 + // . Q . . : off =(-1.5, -1.5)/4 + (1.0, 3.0)/4 + // Static screenspace sample offsets (compute some implicitly): + static const float grid_size = 4.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0,1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5,0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(0.0, 1.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = w1.bgr; + const float3 w3 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0,1.0,1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * (w0 * sample0 + w1 * sample1 + + w2 * sample2 + w3 * sample3); +} + +float3 tex2Daa5x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 5-queens pattern: + // . Q . . . : off =(-2.0, -2.0)/5 + (1.0, 0.0)/5 + // . . . . Q : off =(-2.0, -2.0)/5 + (4.0, 1.0)/5 + // . . Q . . : off =(-2.0, -2.0)/5 + (2.0, 2.0)/5 + // Q . . . . : off =(-2.0, -2.0)/5 + (0.0, 3.0)/5 + // . . . Q . : off =(-2.0, -2.0)/5 + (3.0, 4.0)/5 + // Static screenspace sample offsets (compute some implicitly): + static const float grid_size = 5.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(1.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(4.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(2.0, 2.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = w1.bgr; + const float3 w4 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 w_sum_inv = float3(1.0)/(w0 + w1 + w2 + w3 + w4); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * (w0 * sample0 + w1 * sample1 + + w2 * sample2 + w3 * sample3 + w4 * sample4); +} + +float3 tex2Daa6x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 6-queens pattern with a stronger horizontal + // than vertical slant: + // . . . . Q . : off =(-2.5, -2.5)/6 + (4.0, 0.0)/6 + // . . Q . . . : off =(-2.5, -2.5)/6 + (2.0, 1.0)/6 + // Q . . . . . : off =(-2.5, -2.5)/6 + (0.0, 2.0)/6 + // . . . . . Q : off =(-2.5, -2.5)/6 + (5.0, 3.0)/6 + // . . . Q . . : off =(-2.5, -2.5)/6 + (3.0, 4.0)/6 + // . Q . . . . : off =(-2.5, -2.5)/6 + (1.0, 5.0)/6 + // Static screenspace sample offsets (compute some implicitly): + static const float grid_size = 6.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(4.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(2.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(0.0, 2.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = w2.bgr; + const float3 w4 = w1.bgr; + const float3 w5 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * (w0 * sample0 + w1 * sample1 + w2 * sample2 + + w3 * sample3 + w4 * sample4 + w5 * sample5); +} + +float3 tex2Daa7x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 7-queens pattern with a queen in the center: + // . Q . . . . . : off =(-3.0, -3.0)/7 + (1.0, 0.0)/7 + // . . . . Q . . : off =(-3.0, -3.0)/7 + (4.0, 1.0)/7 + // Q . . . . . . : off =(-3.0, -3.0)/7 + (0.0, 2.0)/7 + // . . . Q . . . : off =(-3.0, -3.0)/7 + (3.0, 3.0)/7 + // . . . . . . Q : off =(-3.0, -3.0)/7 + (6.0, 4.0)/7 + // . . Q . . . . : off =(-3.0, -3.0)/7 + (2.0, 5.0)/7 + // . . . . . Q . : off =(-3.0, -3.0)/7 + (5.0, 6.0)/7 + static const float grid_size = 7.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(1.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(4.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(0.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(3.0, 3.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = w2.bgr; + const float3 w5 = w1.bgr; + const float3 w6 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2; + const float3 w_sum = half_sum + half_sum.bgr + w3; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6); +} + +float3 tex2Daa8x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 8-queens pattern. + // . . Q . . . . . : off =(-3.5, -3.5)/8 + (2.0, 0.0)/8 + // . . . . Q . . . : off =(-3.5, -3.5)/8 + (4.0, 1.0)/8 + // . Q . . . . . . : off =(-3.5, -3.5)/8 + (1.0, 2.0)/8 + // . . . . . . . Q : off =(-3.5, -3.5)/8 + (7.0, 3.0)/8 + // Q . . . . . . . : off =(-3.5, -3.5)/8 + (0.0, 4.0)/8 + // . . . . . . Q . : off =(-3.5, -3.5)/8 + (6.0, 5.0)/8 + // . . . Q . . . . : off =(-3.5, -3.5)/8 + (3.0, 6.0)/8 + // . . . . . Q . . : off =(-3.5, -3.5)/8 + (5.0, 7.0)/8 + static const float grid_size = 8.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(4.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(1.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(7.0, 3.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = w3.bgr; + const float3 w5 = w2.bgr; + const float3 w6 = w1.bgr; + const float3 w7 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, and mirror on odd frames if directed: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7); +} + +float3 tex2Daa12x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 12-superqueens pattern where no 3 points are + // exactly collinear. + // . . . Q . . . . . . . . : off =(-5.5, -5.5)/12 + (3.0, 0.0)/12 + // . . . . . . . . . Q . . : off =(-5.5, -5.5)/12 + (9.0, 1.0)/12 + // . . . . . . Q . . . . . : off =(-5.5, -5.5)/12 + (6.0, 2.0)/12 + // . Q . . . . . . . . . . : off =(-5.5, -5.5)/12 + (1.0, 3.0)/12 + // . . . . . . . . . . . Q : off =(-5.5, -5.5)/12 + (11.0, 4.0)/12 + // . . . . Q . . . . . . . : off =(-5.5, -5.5)/12 + (4.0, 5.0)/12 + // . . . . . . . Q . . . . : off =(-5.5, -5.5)/12 + (7.0, 6.0)/12 + // Q . . . . . . . . . . . : off =(-5.5, -5.5)/12 + (0.0, 7.0)/12 + // . . . . . . . . . . Q . : off =(-5.5, -5.5)/12 + (10.0, 8.0)/12 + // . . . . . Q . . . . . . : off =(-5.5, -5.5)/12 + (5.0, 9.0)/12 + // . . Q . . . . . . . . . : off =(-5.5, -5.5)/12 + (2.0, 10.0)/12 + // . . . . . . . . Q . . . : off =(-5.5, -5.5)/12 + (8.0, 11.0)/12 + static const float grid_size = 12.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(3.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(9.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(6.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(1.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(11.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(4.0, 5.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = w5.bgr; + const float3 w7 = w4.bgr; + const float3 w8 = w3.bgr; + const float3 w9 = w2.bgr; + const float3 w10 = w1.bgr; + const float3 w11 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/w_sum; + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11); +} + +float3 tex2Daa16x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 16-superqueens pattern where no 3 points are + // exactly collinear. + // . . Q . . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (2.0, 0.0)/16 + // . . . . . . . . . Q . . . . . . : off =(-7.5, -7.5)/16 + (9.0, 1.0)/16 + // . . . . . . . . . . . . Q . . . : off =(-7.5, -7.5)/16 + (12.0, 2.0)/16 + // . . . . Q . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (4.0, 3.0)/16 + // . . . . . . . . Q . . . . . . . : off =(-7.5, -7.5)/16 + (8.0, 4.0)/16 + // . . . . . . . . . . . . . . Q . : off =(-7.5, -7.5)/16 + (14.0, 5.0)/16 + // Q . . . . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (0.0, 6.0)/16 + // . . . . . . . . . . Q . . . . . : off =(-7.5, -7.5)/16 + (10.0, 7.0)/16 + // . . . . . Q . . . . . . . . . . : off =(-7.5, -7.5)/16 + (5.0, 8.0)/16 + // . . . . . . . . . . . . . . . Q : off =(-7.5, -7.5)/16 + (15.0, 9.0)/16 + // . Q . . . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (1.0, 10.0)/16 + // . . . . . . . Q . . . . . . . . : off =(-7.5, -7.5)/16 + (7.0, 11.0)/16 + // . . . . . . . . . . . Q . . . . : off =(-7.5, -7.5)/16 + (11.0, 12.0)/16 + // . . . Q . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (3.0, 13.0)/16 + // . . . . . . Q . . . . . . . . . : off =(-7.5, -7.5)/16 + (6.0, 14.0)/16 + // . . . . . . . . . . . . . Q . . : off =(-7.5, -7.5)/16 + (13.0, 15.0)/16 + static const float grid_size = 16.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(9.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(12.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(4.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(8.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(14.0, 5.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(0.0, 6.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(10.0, 7.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = w7.bgr; + const float3 w9 = w6.bgr; + const float3 w10 = w5.bgr; + const float3 w11 = w4.bgr; + const float3 w12 = w3.bgr; + const float3 w13 = w2.bgr; + const float3 w14 = w1.bgr; + const float3 w15 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + const float2 uv_offset6 = mul(true_pixel_to_tex_uv, xy_offset6 * frame_sign); + const float2 uv_offset7 = mul(true_pixel_to_tex_uv, xy_offset7 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset6).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset7).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset7).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset6).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15); +} + +float3 tex2Daa20x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 20-superqueens pattern where no 3 points are + // exactly collinear and superqueens have a squared attack radius of 13. + // . . . . . . . Q . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (7.0, 0.0)/20 + // . . . . . . . . . . . . . . . . Q . . . : off =(-9.5, -9.5)/20 + (16.0, 1.0)/20 + // . . . . . . . . . . . Q . . . . . . . . : off =(-9.5, -9.5)/20 + (11.0, 2.0)/20 + // . Q . . . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (1.0, 3.0)/20 + // . . . . . Q . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (5.0, 4.0)/20 + // . . . . . . . . . . . . . . . Q . . . . : off =(-9.5, -9.5)/20 + (15.0, 5.0)/20 + // . . . . . . . . . . Q . . . . . . . . . : off =(-9.5, -9.5)/20 + (10.0, 6.0)/20 + // . . . . . . . . . . . . . . . . . . . Q : off =(-9.5, -9.5)/20 + (19.0, 7.0)/20 + // . . Q . . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (2.0, 8.0)/20 + // . . . . . . Q . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (6.0, 9.0)/20 + // . . . . . . . . . . . . . Q . . . . . . : off =(-9.5, -9.5)/20 + (13.0, 10.0)/20 + // . . . . . . . . . . . . . . . . . Q . . : off =(-9.5, -9.5)/20 + (17.0, 11.0)/20 + // Q . . . . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (0.0, 12.0)/20 + // . . . . . . . . . Q . . . . . . . . . . : off =(-9.5, -9.5)/20 + (9.0, 13.0)/20 + // . . . . Q . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (4.0, 14.0)/20 + // . . . . . . . . . . . . . . Q . . . . . : off =(-9.5, -9.5)/20 + (14.0, 15.0)/20 + // . . . . . . . . . . . . . . . . . . Q . : off =(-9.5, -9.5)/20 + (18.0, 16.0)/20 + // . . . . . . . . Q . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (8.0, 17.0)/20 + // . . . Q . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (3.0, 18.0)/20 + // . . . . . . . . . . . . Q . . . . . . . : off =(-9.5, -9.5)/20 + (12.0, 19.0)/20 + static const float grid_size = 20.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(7.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(16.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(11.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(1.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(5.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(15.0, 5.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(10.0, 6.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(19.0, 7.0) * xy_step; + const float2 xy_offset8 = xy_start_offset + float2(2.0, 8.0) * xy_step; + const float2 xy_offset9 = xy_start_offset + float2(6.0, 9.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = eval_unorm_rgb_weights(xy_offset8, final_axis_importance); + const float3 w9 = eval_unorm_rgb_weights(xy_offset9, final_axis_importance); + const float3 w10 = w9.bgr; + const float3 w11 = w8.bgr; + const float3 w12 = w7.bgr; + const float3 w13 = w6.bgr; + const float3 w14 = w5.bgr; + const float3 w15 = w4.bgr; + const float3 w16 = w3.bgr; + const float3 w17 = w2.bgr; + const float3 w18 = w1.bgr; + const float3 w19 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + const float2 uv_offset6 = mul(true_pixel_to_tex_uv, xy_offset6 * frame_sign); + const float2 uv_offset7 = mul(true_pixel_to_tex_uv, xy_offset7 * frame_sign); + const float2 uv_offset8 = mul(true_pixel_to_tex_uv, xy_offset8 * frame_sign); + const float2 uv_offset9 = mul(true_pixel_to_tex_uv, xy_offset9 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset6).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset7).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset8).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset9).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset9).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset8).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset7).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset6).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample16 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample17 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample18 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample19 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15 + + w16 * sample16 + w17 * sample17 + w18 * sample18 + w19 * sample19); +} + +float3 tex2Daa24x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 24-superqueens pattern where no 3 points are + // exactly collinear and superqueens have a squared attack radius of 13. + // . . . . . . Q . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (6.0, 0.0)/24 + // . . . . . . . . . . . . . . . . Q . . . . . . . : off =(-11.5, -11.5)/24 + (16.0, 1.0)/24 + // . . . . . . . . . . Q . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (10.0, 2.0)/24 + // . . . . . . . . . . . . . . . . . . . . . Q . . : off =(-11.5, -11.5)/24 + (21.0, 3.0)/24 + // . . . . . Q . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (5.0, 4.0)/24 + // . . . . . . . . . . . . . . . Q . . . . . . . . : off =(-11.5, -11.5)/24 + (15.0, 5.0)/24 + // . Q . . . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (1.0, 6.0)/24 + // . . . . . . . . . . . Q . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (11.0, 7.0)/24 + // . . . . . . . . . . . . . . . . . . . Q . . . . : off =(-11.5, -11.5)/24 + (19.0, 8.0)/24 + // . . . . . . . . . . . . . . . . . . . . . . . Q : off =(-11.5, -11.5)/24 + (23.0, 9.0)/24 + // . . . Q . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (3.0, 10.0)/24 + // . . . . . . . . . . . . . . Q . . . . . . . . . : off =(-11.5, -11.5)/24 + (14.0, 11.0)/24 + // . . . . . . . . . Q . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (9.0, 12.0)/24 + // . . . . . . . . . . . . . . . . . . . . Q . . . : off =(-11.5, -11.5)/24 + (20.0, 13.0)/24 + // Q . . . . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (0.0, 14.0)/24 + // . . . . Q . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (4.0, 15.0)/24 + // . . . . . . . . . . . . Q . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (12.0, 16.0)/24 + // . . . . . . . . . . . . . . . . . . . . . . Q . : off =(-11.5, -11.5)/24 + (22.0, 17.0)/24 + // . . . . . . . . Q . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (8.0, 18.0)/24 + // . . . . . . . . . . . . . . . . . . Q . . . . . : off =(-11.5, -11.5)/24 + (18.0, 19.0)/24 + // . . Q . . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (2.0, 20.0)/24 + // . . . . . . . . . . . . . Q . . . . . . . . . . : off =(-11.5, -11.5)/24 + (13.0, 21.0)/24 + // . . . . . . . Q . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (7.0, 22.0)/24 + // . . . . . . . . . . . . . . . . . Q . . . . . . : off =(-11.5, -11.5)/24 + (17.0, 23.0)/24 + static const float grid_size = 24.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(6.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(16.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(10.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(21.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(5.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(15.0, 5.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(1.0, 6.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(11.0, 7.0) * xy_step; + const float2 xy_offset8 = xy_start_offset + float2(19.0, 8.0) * xy_step; + const float2 xy_offset9 = xy_start_offset + float2(23.0, 9.0) * xy_step; + const float2 xy_offset10 = xy_start_offset + float2(3.0, 10.0) * xy_step; + const float2 xy_offset11 = xy_start_offset + float2(14.0, 11.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = eval_unorm_rgb_weights(xy_offset8, final_axis_importance); + const float3 w9 = eval_unorm_rgb_weights(xy_offset9, final_axis_importance); + const float3 w10 = eval_unorm_rgb_weights(xy_offset10, final_axis_importance); + const float3 w11 = eval_unorm_rgb_weights(xy_offset11, final_axis_importance); + const float3 w12 = w11.bgr; + const float3 w13 = w10.bgr; + const float3 w14 = w9.bgr; + const float3 w15 = w8.bgr; + const float3 w16 = w7.bgr; + const float3 w17 = w6.bgr; + const float3 w18 = w5.bgr; + const float3 w19 = w4.bgr; + const float3 w20 = w3.bgr; + const float3 w21 = w2.bgr; + const float3 w22 = w1.bgr; + const float3 w23 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + + w5 + w6 + w7 + w8 + w9 + w10 + w11; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + const float2 uv_offset6 = mul(true_pixel_to_tex_uv, xy_offset6 * frame_sign); + const float2 uv_offset7 = mul(true_pixel_to_tex_uv, xy_offset7 * frame_sign); + const float2 uv_offset8 = mul(true_pixel_to_tex_uv, xy_offset8 * frame_sign); + const float2 uv_offset9 = mul(true_pixel_to_tex_uv, xy_offset9 * frame_sign); + const float2 uv_offset10 = mul(true_pixel_to_tex_uv, xy_offset10 * frame_sign); + const float2 uv_offset11 = mul(true_pixel_to_tex_uv, xy_offset11 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset6).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset7).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset8).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset9).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset10).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset11).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset11).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset10).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset9).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset8).rgb; + const float3 sample16 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset7).rgb; + const float3 sample17 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset6).rgb; + const float3 sample18 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample19 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample20 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample21 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample22 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample23 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15 + + w16 * sample16 + w17 * sample17 + w18 * sample18 + w19 * sample19 + + w20 * sample20 + w21 * sample21 + w22 * sample22 + w23 * sample23); +} + +float3 tex2Daa_debug_16x_regular(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Sample on a regular 4x4 grid. This is mainly for testing. + static const float grid_size = 4.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample: + const float2 xy_offset0 = xy_start_offset + float2(0.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(1.0, 0.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(3.0, 0.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(0.0, 1.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(1.0, 1.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(2.0, 1.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(3.0, 1.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + // (We can't exploit vertical or horizontal symmetry due to uncertain + // subpixel offsets. We could fix that by rotating xy offsets with the + // subpixel structure, but...no.) + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = w7.bgr; + const float3 w9 = w6.bgr; + const float3 w10 = w5.bgr; + const float3 w11 = w4.bgr; + const float3 w12 = w3.bgr; + const float3 w13 = w2.bgr; + const float3 w14 = w1.bgr; + const float3 w15 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, taking advantage of row alignment: + const float2 uv_step_x = mul(true_pixel_to_tex_uv, float2(xy_step.x, 0.0)); + const float2 uv_step_y = mul(true_pixel_to_tex_uv, float2(0.0, xy_step.y)); + const float2 uv_offset0 = -1.5 * (uv_step_x + uv_step_y); + const float2 sample0_uv = tex_uv + uv_offset0; + const float2 sample4_uv = sample0_uv + uv_step_y; + const float2 sample8_uv = sample0_uv + uv_step_y * 2.0; + const float2 sample12_uv = sample0_uv + uv_step_y * 3.0; + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, sample0_uv).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, sample0_uv + uv_step_x).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, sample0_uv + uv_step_x * 2.0).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, sample0_uv + uv_step_x * 3.0).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, sample4_uv + uv_step_x).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, sample4_uv + uv_step_x * 2.0).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, sample4_uv + uv_step_x * 3.0).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, sample8_uv).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, sample8_uv + uv_step_x).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, sample8_uv + uv_step_x * 2.0).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, sample8_uv + uv_step_x * 3.0).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, sample12_uv).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, sample12_uv + uv_step_x).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, sample12_uv + uv_step_x * 2.0).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, sample12_uv + uv_step_x * 3.0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15); +} + +float3 tex2Daa_debug_dynamic(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // This function is for testing only: Use an NxN grid with dynamic weights. + static const int grid_size = 8; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float grid_radius_in_samples = (float(grid_size) - 1.0)/2.0; + const float2 filter_space_offset_step = + subpixel_support_diameter/float2(grid_size); + const float2 sample0_filter_space_offset = + -grid_radius_in_samples * filter_space_offset_step; + // Compute xy sample offsets and subpixel weights: + float3 weights[64]; //originally grid_size * grid_size + float3 weight_sum = float3(0.0, 0.0, 0.0); + for(int i = 0; i < grid_size; ++i) + { + for(int j = 0; j < grid_size; ++j) + { + // Weights based on xy distances: + const float2 offset = sample0_filter_space_offset + + float2(j, i) * filter_space_offset_step; + const float3 weight = eval_unorm_rgb_weights(offset, final_axis_importance); + weights[i*grid_size + j] = weight; + weight_sum += weight; + } + } + // Get uv offset vectors along x and y directions: + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + const float2 uv_offset_step_x = mul(true_pixel_to_tex_uv, + float2(filter_space_offset_step.x, 0.0)); + const float2 uv_offset_step_y = mul(true_pixel_to_tex_uv, + float2(0.0, filter_space_offset_step.y)); + // Get a starting sample location: + const float2 sample0_uv_offset = -grid_radius_in_samples * + (uv_offset_step_x + uv_offset_step_y); + const float2 sample0_uv = tex_uv + sample0_uv_offset; + // Load, weight, and sum [linearized] samples: + float3 sum = float3(0.0, 0.0, 0.0); + const float3 weight_sum_inv = float3(1.0)/weight_sum; + for(int i = 0; i < grid_size; ++i) + { + const float2 row_i_first_sample_uv = + sample0_uv + i * uv_offset_step_y; + for(int j = 0; j < grid_size; ++j) + { + const float2 sample_uv = + row_i_first_sample_uv + j * uv_offset_step_x; + sum += weights[i*grid_size + j] * + tex2Daa_tiled_linearize(tex, sample_uv).rgb; + } + } + return sum * weight_sum_inv; +} + + +/////////////////////// ANTIALIASING CODEPATH SELECTION ////////////////////// + +inline float3 tex2Daa(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ +#ifdef DEBUG + return tex2Daa_subpixel_weights_only( + tex, tex_uv, pixel_to_tex_uv); +#else + // Statically switch between antialiasing modes/levels: + return (aa_level < 0.5) ? tex2D_linearize(tex, tex_uv).rgb : + (aa_level < 3.5) ? tex2Daa_subpixel_weights_only( + tex, tex_uv, pixel_to_tex_uv) : + (aa_level < 4.5) ? tex2Daa4x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 5.5) ? tex2Daa5x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 6.5) ? tex2Daa6x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 7.5) ? tex2Daa7x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 11.5) ? tex2Daa8x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 15.5) ? tex2Daa12x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 19.5) ? tex2Daa16x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 23.5) ? tex2Daa20x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 253.5) ? tex2Daa24x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 254.5) ? tex2Daa_debug_16x_regular( + tex, tex_uv, pixel_to_tex_uv, frame) : + tex2Daa_debug_dynamic(tex, tex_uv, pixel_to_tex_uv, frame); +#endif +} + + +#endif // TEX2DANTIALIAS_H + +///////////////////////// END TEX2DANTIALIAS ///////////////////////// + +//#include "geometry-functions.h" + +///////////////////////// BEGIN GEOMETRY-FUNCTIONS ///////////////////////// + +#ifndef GEOMETRY_FUNCTIONS_H +#define GEOMETRY_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// already included elsewhere +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" +//#include "bind-shader-h" + + +//////////////////////////// MACROS AND CONSTANTS //////////////////////////// + +// Curvature-related constants: +#define MAX_POINT_CLOUD_SIZE 9 + + +///////////////////////////// CURVATURE FUNCTIONS ///////////////////////////// + +float2 quadratic_solve(const float a, const float b_over_2, const float c) +{ + // Requires: 1.) a, b, and c are quadratic formula coefficients + // 2.) b_over_2 = b/2.0 (simplifies terms to factor 2 out) + // 3.) b_over_2 must be guaranteed < 0.0 (avoids a branch) + // Returns: Returns float2(first_solution, discriminant), so the caller + // can choose how to handle the "no intersection" case. The + // Kahan or Citardauq formula is used for numerical robustness. + const float discriminant = b_over_2*b_over_2 - a*c; + const float solution0 = c/(-b_over_2 + sqrt(discriminant)); + return float2(solution0, discriminant); +} + +float2 intersect_sphere(const float3 view_vec, const float3 eye_pos_vec) +{ + // Requires: 1.) view_vec and eye_pos_vec are 3D vectors in the sphere's + // local coordinate frame (eye_pos_vec is a position, i.e. + // a vector from the origin to the eye/camera) + // 2.) geom_radius is a global containing the sphere's radius + // Returns: Cast a ray of direction view_vec from eye_pos_vec at a + // sphere of radius geom_radius, and return the distance to + // the first intersection in units of length(view_vec). + // http://wiki.cgsociety.org/index.php/Ray_Sphere_Intersection + // Quadratic formula coefficients (b_over_2 is guaranteed negative): + const float a = dot(view_vec, view_vec); + const float b_over_2 = dot(view_vec, eye_pos_vec); // * 2.0 factored out + const float c = dot(eye_pos_vec, eye_pos_vec) - geom_radius*geom_radius; + return quadratic_solve(a, b_over_2, c); +} + +float2 intersect_cylinder(const float3 view_vec, const float3 eye_pos_vec) +{ + // Requires: 1.) view_vec and eye_pos_vec are 3D vectors in the sphere's + // local coordinate frame (eye_pos_vec is a position, i.e. + // a vector from the origin to the eye/camera) + // 2.) geom_radius is a global containing the cylinder's radius + // Returns: Cast a ray of direction view_vec from eye_pos_vec at a + // cylinder of radius geom_radius, and return the distance to + // the first intersection in units of length(view_vec). The + // derivation of the coefficients is in Christer Ericson's + // Real-Time Collision Detection, p. 195-196, and this version + // uses LaGrange's identity to reduce operations. + // Arbitrary "cylinder top" reference point for an infinite cylinder: + const float3 cylinder_top_vec = float3(0.0, geom_radius, 0.0); + const float3 cylinder_axis_vec = float3(0.0, 1.0, 0.0);//float3(0.0, 2.0*geom_radius, 0.0); + const float3 top_to_eye_vec = eye_pos_vec - cylinder_top_vec; + const float3 axis_x_view = cross(cylinder_axis_vec, view_vec); + const float3 axis_x_top_to_eye = cross(cylinder_axis_vec, top_to_eye_vec); + // Quadratic formula coefficients (b_over_2 is guaranteed negative): + const float a = dot(axis_x_view, axis_x_view); + const float b_over_2 = dot(axis_x_top_to_eye, axis_x_view); + const float c = dot(axis_x_top_to_eye, axis_x_top_to_eye) - + geom_radius*geom_radius;//*dot(cylinder_axis_vec, cylinder_axis_vec); + return quadratic_solve(a, b_over_2, c); +} + +float2 cylinder_xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect) +{ + // Requires: An xyz intersection position on a cylinder. + // Returns: video_uv coords mapped to range [-0.5, 0.5] + // Mapping: Define square_uv.x to be the signed arc length in xz-space, + // and define square_uv.y = -intersection_pos_local.y (+v = -y). + // Start with a numerically robust arc length calculation. + const float angle_from_image_center = atan2(intersection_pos_local.x, + intersection_pos_local.z); + const float signed_arc_len = angle_from_image_center * geom_radius; + // Get a uv-mapping where [-0.5, 0.5] maps to a "square" area, then divide + // by the aspect ratio to stretch the mapping appropriately: + const float2 square_uv = float2(signed_arc_len, -intersection_pos_local.y); + const float2 video_uv = square_uv / geom_aspect; + return video_uv; +} + +float3 cylinder_uv_to_xyz(const float2 video_uv, const float2 geom_aspect) +{ + // Requires: video_uv coords mapped to range [-0.5, 0.5] + // Returns: An xyz intersection position on a cylinder. This is the + // inverse of cylinder_xyz_to_uv(). + // Expand video_uv by the aspect ratio to get proportionate x/y lengths, + // then calculate an xyz position for the cylindrical mapping above. + const float2 square_uv = video_uv * geom_aspect; + const float arc_len = square_uv.x; + const float angle_from_image_center = arc_len / geom_radius; + const float x_pos = sin(angle_from_image_center) * geom_radius; + const float z_pos = cos(angle_from_image_center) * geom_radius; + // Or: z = sqrt(geom_radius**2 - x**2) + // Or: z = geom_radius/sqrt(1.0 + tan(angle)**2), x = z * tan(angle) + const float3 intersection_pos_local = float3(x_pos, -square_uv.y, z_pos); + return intersection_pos_local; +} + +float2 sphere_xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect) +{ + // Requires: An xyz intersection position on a sphere. + // Returns: video_uv coords mapped to range [-0.5, 0.5] + // Mapping: First define square_uv.x/square_uv.y == + // intersection_pos_local.x/intersection_pos_local.y. Then, + // length(square_uv) is the arc length from the image center + // at (0.0, 0.0, geom_radius) along the tangent great circle. + // Credit for this mapping goes to cgwg: I never managed to + // understand his code, but he told me his mapping was based on + // great circle distances when I asked him about it, which + // informed this very similar (almost identical) mapping. + // Start with a numerically robust arc length calculation between the ray- + // sphere intersection point and the image center using a method posted by + // Roger Stafford on comp.soft-sys.matlab: + // https://groups.google.com/d/msg/comp.soft-sys.matlab/zNbUui3bjcA/c0HV_bHSx9cJ + const float3 image_center_pos_local = float3(0.0, 0.0, geom_radius); + const float cp_len = + length(cross(intersection_pos_local, image_center_pos_local)); + const float dp = dot(intersection_pos_local, image_center_pos_local); + const float angle_from_image_center = atan2(cp_len, dp); + const float arc_len = angle_from_image_center * geom_radius; + // Get a uv-mapping where [-0.5, 0.5] maps to a "square" area, then divide + // by the aspect ratio to stretch the mapping appropriately: + const float2 square_uv_unit = normalize(float2(intersection_pos_local.x, + -intersection_pos_local.y)); + const float2 square_uv = arc_len * square_uv_unit; + const float2 video_uv = square_uv / geom_aspect; + return video_uv; +} + +float3 sphere_uv_to_xyz(const float2 video_uv, const float2 geom_aspect) +{ + // Requires: video_uv coords mapped to range [-0.5, 0.5] + // Returns: An xyz intersection position on a sphere. This is the + // inverse of sphere_xyz_to_uv(). + // Expand video_uv by the aspect ratio to get proportionate x/y lengths, + // then calculate an xyz position for the spherical mapping above. + const float2 square_uv = video_uv * geom_aspect; + // Using length or sqrt here butchers the framerate on my 8800GTS if + // this function is called too many times, and so does taking the max + // component of square_uv/square_uv_unit (program length threshold?). + //float arc_len = length(square_uv); + const float2 square_uv_unit = normalize(square_uv); + const float arc_len = square_uv.y/square_uv_unit.y; + const float angle_from_image_center = arc_len / geom_radius; + const float xy_dist_from_sphere_center = + sin(angle_from_image_center) * geom_radius; + //float2 xy_pos = xy_dist_from_sphere_center * (square_uv/FIX_ZERO(arc_len)); + const float2 xy_pos = xy_dist_from_sphere_center * square_uv_unit; + const float z_pos = cos(angle_from_image_center) * geom_radius; + const float3 intersection_pos_local = float3(xy_pos.x, -xy_pos.y, z_pos); + return intersection_pos_local; +} + +float2 sphere_alt_xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect) +{ + // Requires: An xyz intersection position on a cylinder. + // Returns: video_uv coords mapped to range [-0.5, 0.5] + // Mapping: Define square_uv.x to be the signed arc length in xz-space, + // and define square_uv.y == signed arc length in yz-space. + // See cylinder_xyz_to_uv() for implementation details (very similar). + const float2 angle_from_image_center = atan2( + float2(intersection_pos_local.x, -intersection_pos_local.y), + intersection_pos_local.zz); + const float2 signed_arc_len = angle_from_image_center * geom_radius; + const float2 video_uv = signed_arc_len / geom_aspect; + return video_uv; +} + +float3 sphere_alt_uv_to_xyz(const float2 video_uv, const float2 geom_aspect) +{ + // Requires: video_uv coords mapped to range [-0.5, 0.5] + // Returns: An xyz intersection position on a sphere. This is the + // inverse of sphere_alt_xyz_to_uv(). + // See cylinder_uv_to_xyz() for implementation details (very similar). + const float2 square_uv = video_uv * geom_aspect; + const float2 arc_len = square_uv; + const float2 angle_from_image_center = arc_len / geom_radius; + const float2 xy_pos = sin(angle_from_image_center) * geom_radius; + const float z_pos = sqrt(geom_radius*geom_radius - dot(xy_pos, xy_pos)); + return float3(xy_pos.x, -xy_pos.y, z_pos); +} + +inline float2 intersect(const float3 view_vec_local, const float3 eye_pos_local, + const float geom_mode) +{ + return geom_mode < 2.5 ? intersect_sphere(view_vec_local, eye_pos_local) : + intersect_cylinder(view_vec_local, eye_pos_local); +} + +inline float2 xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect, const float geom_mode) +{ + return geom_mode < 1.5 ? + sphere_xyz_to_uv(intersection_pos_local, geom_aspect) : + geom_mode < 2.5 ? + sphere_alt_xyz_to_uv(intersection_pos_local, geom_aspect) : + cylinder_xyz_to_uv(intersection_pos_local, geom_aspect); +} + +inline float3 uv_to_xyz(const float2 uv, const float2 geom_aspect, + const float geom_mode) +{ + return geom_mode < 1.5 ? sphere_uv_to_xyz(uv, geom_aspect) : + geom_mode < 2.5 ? sphere_alt_uv_to_xyz(uv, geom_aspect) : + cylinder_uv_to_xyz(uv, geom_aspect); +} + +float2 view_vec_to_uv(const float3 view_vec_local, const float3 eye_pos_local, + const float2 geom_aspect, const float geom_mode, out float3 intersection_pos) +{ + // Get the intersection point on the primitive, given an eye position + // and view vector already in its local coordinate frame: + const float2 intersect_dist_and_discriminant = intersect(view_vec_local, + eye_pos_local, geom_mode); + const float3 intersection_pos_local = eye_pos_local + + view_vec_local * intersect_dist_and_discriminant.x; + // Save the intersection position to an output parameter: + intersection_pos = intersection_pos_local; + // Transform into uv coords, but give out-of-range coords if the + // view ray doesn't intersect the primitive in the first place: + return intersect_dist_and_discriminant.y > 0.005 ? + xyz_to_uv(intersection_pos_local, geom_aspect, geom_mode) : float2(1.0); +} + +float3 get_ideal_global_eye_pos_for_points(float3 eye_pos, + const float2 geom_aspect, const float3 global_coords[MAX_POINT_CLOUD_SIZE], + const int num_points) +{ + // Requires: Parameters: + // 1.) Starting eye_pos is a global 3D position at which the + // camera contains all points in global_coords[] in its FOV + // 2.) geom_aspect = get_aspect_vector( + // output_size.x / output_size.y); + // 3.) global_coords is a point cloud containing global xyz + // coords of extreme points on the simulated CRT screen. + // Globals: + // 1.) geom_view_dist must be > 0.0. It controls the "near + // plane" used to interpret flat_video_uv as a view + // vector, which controls the field of view (FOV). + // Eyespace coordinate frame: +x = right, +y = up, +z = back + // Returns: Return an eye position at which the point cloud spans as + // much of the screen as possible (given the FOV controlled by + // geom_view_dist) without being cropped or sheared. + // Algorithm: + // 1.) Move the eye laterally to a point which attempts to maximize the + // the amount we can move forward without clipping the CRT screen. + // 2.) Move forward by as much as possible without clipping the CRT. + // Get the allowed movement range by solving for the eye_pos offsets + // that result in each point being projected to a screen edge/corner in + // pseudo-normalized device coords (where xy ranges from [-0.5, 0.5] + // and z = eyespace z): + // pndc_coord = float3(float2(eyespace_xyz.x, -eyespace_xyz.y)* + // geom_view_dist / (geom_aspect * -eyespace_xyz.z), eyespace_xyz.z); + // Notes: + // The field of view is controlled by geom_view_dist's magnitude relative to + // the view vector's x and y components: + // view_vec.xy ranges from [-0.5, 0.5] * geom_aspect + // view_vec.z = -geom_view_dist + // But for the purposes of perspective divide, it should be considered: + // view_vec.xy ranges from [-0.5, 0.5] * geom_aspect / geom_view_dist + // view_vec.z = -1.0 + static const int max_centering_iters = 1; // Keep for easy testing. + for(int iter = 0; iter < max_centering_iters; iter++) + { + // 0.) Get the eyespace coordinates of our point cloud: + float3 eyespace_coords[MAX_POINT_CLOUD_SIZE]; + for(int i = 0; i < num_points; i++) + { + eyespace_coords[i] = global_coords[i] - eye_pos; + } + // 1a.)For each point, find out how far we can move eye_pos in each + // lateral direction without the point clipping the frustum. + // Eyespace +y = up, screenspace +y = down, so flip y after + // applying the eyespace offset (on the way to "clip space"). + // Solve for two offsets per point based on: + // (eyespace_xyz.xy - offset_dr) * float2(1.0, -1.0) * + // geom_view_dist / (geom_aspect * -eyespace_xyz.z) = float2(-0.5) + // (eyespace_xyz.xy - offset_dr) * float2(1.0, -1.0) * + // geom_view_dist / (geom_aspect * -eyespace_xyz.z) = float2(0.5) + // offset_ul and offset_dr represent the farthest we can move the + // eye_pos up-left and down-right. Save the min of all offset_dr's + // and the max of all offset_ul's (since it's negative). + float abs_radius = abs(geom_radius); // In case anyone gets ideas. ;) + float2 offset_dr_min = float2(10.0 * abs_radius, 10.0 * abs_radius); + float2 offset_ul_max = float2(-10.0 * abs_radius, -10.0 * abs_radius); + for(int i = 0; i < num_points; i++) + { + static const float2 flipy = float2(1.0, -1.0); + float3 eyespace_xyz = eyespace_coords[i]; + float2 offset_dr = eyespace_xyz.xy - float2(-0.5) * + (geom_aspect * -eyespace_xyz.z) / (geom_view_dist * flipy); + float2 offset_ul = eyespace_xyz.xy - float2(0.5) * + (geom_aspect * -eyespace_xyz.z) / (geom_view_dist * flipy); + offset_dr_min = min(offset_dr_min, offset_dr); + offset_ul_max = max(offset_ul_max, offset_ul); + } + // 1b.)Update eye_pos: Adding the average of offset_ul_max and + // offset_dr_min gives it equal leeway on the top vs. bottom + // and left vs. right. Recalculate eyespace_coords accordingly. + float2 center_offset = 0.5 * (offset_ul_max + offset_dr_min); + eye_pos.xy += center_offset; + for(int i = 0; i < num_points; i++) + { + eyespace_coords[i] = global_coords[i] - eye_pos; + } + // 2a.)For each point, find out how far we can move eye_pos forward + // without the point clipping the frustum. Flip the y + // direction in advance (matters for a later step, not here). + // Solve for four offsets per point based on: + // eyespace_xyz_flipy.x * geom_view_dist / + // (geom_aspect.x * (offset_z - eyespace_xyz_flipy.z)) =-0.5 + // eyespace_xyz_flipy.y * geom_view_dist / + // (geom_aspect.y * (offset_z - eyespace_xyz_flipy.z)) =-0.5 + // eyespace_xyz_flipy.x * geom_view_dist / + // (geom_aspect.x * (offset_z - eyespace_xyz_flipy.z)) = 0.5 + // eyespace_xyz_flipy.y * geom_view_dist / + // (geom_aspect.y * (offset_z - eyespace_xyz_flipy.z)) = 0.5 + // We'll vectorize the actual computation. Take the maximum of + // these four for a single offset, and continue taking the max + // for every point (use max because offset.z is negative). + float offset_z_max = -10.0 * geom_radius * geom_view_dist; + for(int i = 0; i < num_points; i++) + { + float3 eyespace_xyz_flipy = eyespace_coords[i] * + float3(1.0, -1.0, 1.0); + float4 offset_zzzz = eyespace_xyz_flipy.zzzz + + (eyespace_xyz_flipy.xyxy * geom_view_dist) / + (float4(-0.5, -0.5, 0.5, 0.5) * float4(geom_aspect, geom_aspect)); + // Ignore offsets that push positive x/y values to opposite + // boundaries, and vice versa, and don't let the camera move + // past a point in the dead center of the screen: + offset_z_max = (eyespace_xyz_flipy.x < 0.0) ? + max(offset_z_max, offset_zzzz.x) : offset_z_max; + offset_z_max = (eyespace_xyz_flipy.y < 0.0) ? + max(offset_z_max, offset_zzzz.y) : offset_z_max; + offset_z_max = (eyespace_xyz_flipy.x > 0.0) ? + max(offset_z_max, offset_zzzz.z) : offset_z_max; + offset_z_max = (eyespace_xyz_flipy.y > 0.0) ? + max(offset_z_max, offset_zzzz.w) : offset_z_max; + offset_z_max = max(offset_z_max, eyespace_xyz_flipy.z); + } + // 2b.)Update eye_pos: Add the maximum (smallest negative) z offset. + eye_pos.z += offset_z_max; + } + return eye_pos; +} + +float3 get_ideal_global_eye_pos(const float3x3 local_to_global, + const float2 geom_aspect, const float geom_mode) +{ + // Start with an initial eye_pos that includes the entire primitive + // (sphere or cylinder) in its field-of-view: + const float3 high_view = float3(0.0, geom_aspect.y, -geom_view_dist); + const float3 low_view = high_view * float3(1.0, -1.0, 1.0); + const float len_sq = dot(high_view, high_view); + const float fov = abs(acos(dot(high_view, low_view)/len_sq)); + // Trigonometry/similar triangles say distance = geom_radius/sin(fov/2): + const float eye_z_spherical = geom_radius/sin(fov*0.5); + const float3 eye_pos = geom_mode < 2.5 ? + float3(0.0, 0.0, eye_z_spherical) : + float3(0.0, 0.0, max(geom_view_dist, eye_z_spherical)); + + // Get global xyz coords of extreme sample points on the simulated CRT + // screen. Start with the center, edge centers, and corners of the + // video image. We can't ignore backfacing points: They're occluded + // by closer points on the primitive, but they may NOT be occluded by + // the convex hull of the remaining samples (i.e. the remaining convex + // hull might not envelope points that do occlude a back-facing point.) + static const int num_points = MAX_POINT_CLOUD_SIZE; + float3 global_coords[MAX_POINT_CLOUD_SIZE]; + global_coords[0] = mul(local_to_global, uv_to_xyz(float2(0.0, 0.0), geom_aspect, geom_mode)); + global_coords[1] = mul(local_to_global, uv_to_xyz(float2(0.0, -0.5), geom_aspect, geom_mode)); + global_coords[2] = mul(local_to_global, uv_to_xyz(float2(0.0, 0.5), geom_aspect, geom_mode)); + global_coords[3] = mul(local_to_global, uv_to_xyz(float2(-0.5, 0.0), geom_aspect, geom_mode)); + global_coords[4] = mul(local_to_global, uv_to_xyz(float2(0.5, 0.0), geom_aspect, geom_mode)); + global_coords[5] = mul(local_to_global, uv_to_xyz(float2(-0.5, -0.5), geom_aspect, geom_mode)); + global_coords[6] = mul(local_to_global, uv_to_xyz(float2(0.5, -0.5), geom_aspect, geom_mode)); + global_coords[7] = mul(local_to_global, uv_to_xyz(float2(-0.5, 0.5), geom_aspect, geom_mode)); + global_coords[8] = mul(local_to_global, uv_to_xyz(float2(0.5, 0.5), geom_aspect, geom_mode)); + // Adding more inner image points could help in extreme cases, but too many + // points will kille the framerate. For safety, default to the initial + // eye_pos if any z coords are negative: + float num_negative_z_coords = 0.0; + for(int i = 0; i < num_points; i++) + { + num_negative_z_coords += float(global_coords[0].z < 0.0); + } + // Outsource the optimized eye_pos calculation: + return num_negative_z_coords > 0.5 ? eye_pos : + get_ideal_global_eye_pos_for_points(eye_pos, geom_aspect, + global_coords, num_points); +} + +float3x3 get_pixel_to_object_matrix(const float3x3 global_to_local, + const float3 eye_pos_local, const float3 view_vec_global, + const float3 intersection_pos_local, const float3 normal, + const float2 output_size_inv) +{ + // Requires: See get_curved_video_uv_coords_and_tangent_matrix for + // descriptions of each parameter. + // Returns: Return a transformation matrix from 2D pixel-space vectors + // (where (+1.0, +1.0) is a vector to one pixel down-right, + // i.e. same directionality as uv texels) to 3D object-space + // vectors in the CRT's local coordinate frame (right-handed) + // ***which are tangent to the CRT surface at the intersection + // position.*** (Basically, we want to convert pixel-space + // vectors to 3D vectors along the CRT's surface, for later + // conversion to uv vectors.) + // Shorthand inputs: + const float3 pos = intersection_pos_local; + const float3 eye_pos = eye_pos_local; + // Get a piecewise-linear matrix transforming from "pixelspace" offset + // vectors (1.0 = one pixel) to object space vectors in the tangent + // plane (faster than finding 3 view-object intersections). + // 1.) Get the local view vecs for the pixels to the right and down: + const float3 view_vec_right_global = view_vec_global + + float3(output_size_inv.x, 0.0, 0.0); + const float3 view_vec_down_global = view_vec_global + + float3(0.0, -output_size_inv.y, 0.0); + const float3 view_vec_right_local = + mul(global_to_local, view_vec_right_global); + const float3 view_vec_down_local = + mul(global_to_local, view_vec_down_global); + // 2.) Using the true intersection point, intersect the neighboring + // view vectors with the tangent plane: + const float3 intersection_vec_dot_normal = float3(dot(pos - eye_pos, normal), dot(pos - eye_pos, normal), dot(pos - eye_pos, normal)); + const float3 right_pos = eye_pos + (intersection_vec_dot_normal / + dot(view_vec_right_local, normal))*view_vec_right_local; + const float3 down_pos = eye_pos + (intersection_vec_dot_normal / + dot(view_vec_down_local, normal))*view_vec_down_local; + // 3.) Subtract the original intersection pos from its neighbors; the + // resulting vectors are object-space vectors tangent to the plane. + // These vectors are the object-space transformations of (1.0, 0.0) + // and (0.0, 1.0) pixel offsets, so they form the first two basis + // vectors of a pixelspace to object space transformation. This + // transformation is 2D to 3D, so use (0, 0, 0) for the third vector. + const float3 object_right_vec = right_pos - pos; + const float3 object_down_vec = down_pos - pos; + const float3x3 pixel_to_object = float3x3( + object_right_vec.x, object_down_vec.x, 0.0, + object_right_vec.y, object_down_vec.y, 0.0, + object_right_vec.z, object_down_vec.z, 0.0); + return pixel_to_object; +} + +float3x3 get_object_to_tangent_matrix(const float3 intersection_pos_local, + const float3 normal, const float2 geom_aspect, const float geom_mode) +{ + // Requires: See get_curved_video_uv_coords_and_tangent_matrix for + // descriptions of each parameter. + // Returns: Return a transformation matrix from 3D object-space vectors + // in the CRT's local coordinate frame (right-handed, +y = up) + // to 2D video_uv vectors (+v = down). + // Description: + // The TBN matrix formed by the [tangent, bitangent, normal] basis + // vectors transforms ordinary vectors from tangent->object space. + // The cotangent matrix formed by the [cotangent, cobitangent, normal] + // basis vectors transforms normal vectors (covectors) from + // tangent->object space. It's the inverse-transpose of the TBN matrix. + // We want the inverse of the TBN matrix (transpose of the cotangent + // matrix), which transforms ordinary vectors from object->tangent space. + // Start by calculating the relevant basis vectors in accordance with + // Christian Schüler's blog post "Followup: Normal Mapping Without + // Precomputed Tangents": http://www.thetenthplanet.de/archives/1180 + // With our particular uv mapping, the scale of the u and v directions + // is determined entirely by the aspect ratio for cylindrical and ordinary + // spherical mappings, and so tangent and bitangent lengths are also + // determined by it (the alternate mapping is more complex). Therefore, we + // must ensure appropriate cotangent and cobitangent lengths as well. + // Base these off the uv<=>xyz mappings for each primitive. + const float3 pos = intersection_pos_local; + static const float3 x_vec = float3(1.0, 0.0, 0.0); + static const float3 y_vec = float3(0.0, 1.0, 0.0); + // The tangent and bitangent vectors correspond with increasing u and v, + // respectively. Mathematically we'd base the cotangent/cobitangent on + // those, but we'll compute the cotangent/cobitangent directly when we can. + float3 cotangent_unscaled, cobitangent_unscaled; + // geom_mode should be constant-folded without RUNTIME_GEOMETRY_MODE. + if(geom_mode < 1.5) + { + // Sphere: + // tangent = normalize(cross(normal, cross(x_vec, pos))) * geom_aspect.x + // bitangent = normalize(cross(cross(y_vec, pos), normal)) * geom_aspect.y + // inv_determinant = 1.0/length(cross(bitangent, tangent)) + // cotangent = cross(normal, bitangent) * inv_determinant + // == normalize(cross(y_vec, pos)) * geom_aspect.y * inv_determinant + // cobitangent = cross(tangent, normal) * inv_determinant + // == normalize(cross(x_vec, pos)) * geom_aspect.x * inv_determinant + // Simplified (scale by inv_determinant below): + cotangent_unscaled = normalize(cross(y_vec, pos)) * geom_aspect.y; + cobitangent_unscaled = normalize(cross(x_vec, pos)) * geom_aspect.x; + } + else if(geom_mode < 2.5) + { + // Sphere, alternate mapping: + // This mapping works a bit like the cylindrical mapping in two + // directions, which makes the lengths and directions more complex. + // Unfortunately, I can't find much of a shortcut: + const float3 tangent = normalize( + cross(y_vec, float3(pos.x, 0.0, pos.z))) * geom_aspect.x; + const float3 bitangent = normalize( + cross(x_vec, float3(0.0, pos.yz))) * geom_aspect.y; + cotangent_unscaled = cross(normal, bitangent); + cobitangent_unscaled = cross(tangent, normal); + } + else + { + // Cylinder: + // tangent = normalize(cross(y_vec, normal)) * geom_aspect.x; + // bitangent = float3(0.0, -geom_aspect.y, 0.0); + // inv_determinant = 1.0/length(cross(bitangent, tangent)) + // cotangent = cross(normal, bitangent) * inv_determinant + // == normalize(cross(y_vec, pos)) * geom_aspect.y * inv_determinant + // cobitangent = cross(tangent, normal) * inv_determinant + // == float3(0.0, -geom_aspect.x, 0.0) * inv_determinant + cotangent_unscaled = cross(y_vec, normal) * geom_aspect.y; + cobitangent_unscaled = float3(0.0, -geom_aspect.x, 0.0); + } + const float3 computed_normal = + cross(cobitangent_unscaled, cotangent_unscaled); + const float inv_determinant = rsqrt(dot(computed_normal, computed_normal)); + const float3 cotangent = cotangent_unscaled * inv_determinant; + const float3 cobitangent = cobitangent_unscaled * inv_determinant; + // The [cotangent, cobitangent, normal] column vecs form the cotangent + // frame, i.e. the inverse-transpose TBN matrix. Get its transpose: + const float3x3 object_to_tangent = float3x3(cotangent, cobitangent, normal); + return object_to_tangent; +} + +float2 get_curved_video_uv_coords_and_tangent_matrix( + const float2 flat_video_uv, const float3 eye_pos_local, + const float2 output_size_inv, const float2 geom_aspect, + const float geom_mode, const float3x3 global_to_local, + out float2x2 pixel_to_tangent_video_uv) +{ + // Requires: Parameters: + // 1.) flat_video_uv coords are in range [0.0, 1.0], where + // (0.0, 0.0) is the top-left corner of the screen and + // (1.0, 1.0) is the bottom-right corner. + // 2.) eye_pos_local is the 3D camera position in the simulated + // CRT's local coordinate frame. For best results, it must + // be computed based on the same geom_view_dist used here. + // 3.) output_size_inv = float2(1.0)/output_size + // 4.) geom_aspect = get_aspect_vector( + // output_size.x / output_size.y); + // 5.) geom_mode is a static or runtime mode setting: + // 0 = off, 1 = sphere, 2 = sphere alt., 3 = cylinder + // 6.) global_to_local is a 3x3 matrix transforming (ordinary) + // worldspace vectors to the CRT's local coordinate frame + // Globals: + // 1.) geom_view_dist must be > 0.0. It controls the "near + // plane" used to interpret flat_video_uv as a view + // vector, which controls the field of view (FOV). + // Returns: Return final uv coords in [0.0, 1.0], and return a pixel- + // space to video_uv tangent-space matrix in the out parameter. + // (This matrix assumes pixel-space +y = down, like +v = down.) + // We'll transform flat_video_uv into a view vector, project + // the view vector from the camera/eye, intersect with a sphere + // or cylinder representing the simulated CRT, and convert the + // intersection position into final uv coords and a local + // transformation matrix. + // First get the 3D view vector (geom_aspect and geom_view_dist are globals): + // 1.) Center uv around (0.0, 0.0) and make (-0.5, -0.5) and (0.5, 0.5) + // correspond to the top-left/bottom-right output screen corners. + // 2.) Multiply by geom_aspect to preemptively "undo" Retroarch's screen- + // space 2D aspect correction. We'll reapply it in uv-space. + // 3.) (x, y) = (u, -v), because +v is down in 2D screenspace, but +y + // is up in 3D worldspace (enforce a right-handed system). + // 4.) The view vector z controls the "near plane" distance and FOV. + // For the effect of "looking through a window" at a CRT, it should be + // set equal to the user's distance from their physical screen, in + // units of the viewport's physical diagonal size. + const float2 view_uv = (flat_video_uv - float2(0.5)) * geom_aspect; + const float3 view_vec_global = + float3(view_uv.x, -view_uv.y, -geom_view_dist); + // Transform the view vector into the CRT's local coordinate frame, convert + // to video_uv coords, and get the local 3D intersection position: + const float3 view_vec_local = mul(global_to_local, view_vec_global); + float3 pos; + const float2 centered_uv = view_vec_to_uv( + view_vec_local, eye_pos_local, geom_aspect, geom_mode, pos); + const float2 video_uv = centered_uv + float2(0.5); + // Get a pixel-to-tangent-video-uv matrix. The caller could deal with + // all but one of these cases, but that would be more complicated. + #ifdef DRIVERS_ALLOW_DERIVATIVES + // Derivatives obtain a matrix very fast, but the direction of pixel- + // space +y seems to depend on the pass. Enforce the correct direction + // on a best-effort basis (but it shouldn't matter for antialiasing). + const float2 duv_dx = ddx(video_uv); + const float2 duv_dy = ddy(video_uv); + #ifdef LAST_PASS + pixel_to_tangent_video_uv = float2x2( + duv_dx.x, duv_dy.x, + -duv_dx.y, -duv_dy.y); + #else + pixel_to_tangent_video_uv = float2x2( + duv_dx.x, duv_dy.x, + duv_dx.y, duv_dy.y); + #endif + #else + // Manually define a transformation matrix. We'll assume pixel-space + // +y = down, just like +v = down. + if(geom_force_correct_tangent_matrix) + { + // Get the surface normal based on the local intersection position: + const float3 normal_base = geom_mode < 2.5 ? pos : + float3(pos.x, 0.0, pos.z); + const float3 normal = normalize(normal_base); + // Get pixel-to-object and object-to-tangent matrices and combine + // them into a 2x2 pixel-to-tangent matrix for video_uv offsets: + const float3x3 pixel_to_object = get_pixel_to_object_matrix( + global_to_local, eye_pos_local, view_vec_global, pos, normal, + output_size_inv); + const float3x3 object_to_tangent = get_object_to_tangent_matrix( + pos, normal, geom_aspect, geom_mode); + const float3x3 pixel_to_tangent3x3 = + mul(object_to_tangent, pixel_to_object); + pixel_to_tangent_video_uv = float2x2( + pixel_to_tangent3x3[0][0], pixel_to_tangent3x3[0][1], pixel_to_tangent3x3[1][0], pixel_to_tangent3x3[1][1]);//._m00_m01_m10_m11); //TODO/FIXME: needs to correct for column-major?? + } + else + { + // Ignore curvature, and just consider flat scaling. The + // difference is only apparent with strong curvature: + pixel_to_tangent_video_uv = float2x2( + output_size_inv.x, 0.0, 0.0, output_size_inv.y); + } + #endif + return video_uv; +} + +float get_border_dim_factor(const float2 video_uv, const float2 geom_aspect) +{ + // COPYRIGHT NOTE FOR THIS FUNCTION: + // Copyright (C) 2010-2012 cgwg, 2014 TroggleMonkey + // This function uses an algorithm first coded in several of cgwg's GPL- + // licensed lines in crt-geom-curved.cg and its ancestors. The line + // between algorithm and code is nearly indistinguishable here, so it's + // unclear whether I could even release this project under a non-GPL + // license with this function included. + + // Calculate border_dim_factor from the proximity to uv-space image + // borders; geom_aspect/border_size/border/darkness/border_compress are globals: + const float2 edge_dists = min(video_uv, float2(1.0) - video_uv) * + geom_aspect; + const float2 border_penetration = + max(float2(border_size) - edge_dists, float2(0.0)); + const float penetration_ratio = length(border_penetration)/border_size; + const float border_escape_ratio = max(1.0 - penetration_ratio, 0.0); + const float border_dim_factor = + pow(border_escape_ratio, border_darkness) * max(1.0, border_compress); + return min(border_dim_factor, 1.0); +} + + + +#endif // GEOMETRY_FUNCTIONS_H + +///////////////////////// END GEOMETRY-FUNCTIONS ///////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +float2x2 mul_scale(float2 scale, float2x2 matrix) +{ + //float2x2 scale_matrix = float2x2(scale.x, 0.0, 0.0, scale.y); + //return mul(scale_matrix, matrix); + float4 intermed = float4(matrix[0][0],matrix[0][1],matrix[1][0],matrix[1][1]) * scale.xxyy; + return float2x2(intermed.x, intermed.y, intermed.z, intermed.w); +} + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + // Localize some parameters: + const float2 geom_aspect = geom_aspect_and_overscan.xy; + const float2 geom_overscan = geom_aspect_and_overscan.zw; + const float2 video_size_inv = video_and_texture_size_inv.xy; + const float2 texture_size_inv = video_and_texture_size_inv.zw; + //const float2 output_size_inv = output_size_inv; + #ifdef RUNTIME_GEOMETRY_TILT + const float3x3 global_to_local = float3x3(global_to_local_row0, + global_to_local_row1, global_to_local_row2); + #else + static const float3x3 global_to_local = geom_global_to_local_static; + #endif + #ifdef RUNTIME_GEOMETRY_MODE + const float geom_mode = geom_mode_runtime; + #else + static const float geom_mode = geom_mode_static; + #endif + + // Get flat and curved texture coords for the current fragment point sample + // and a pixel_to_tangent_video_uv matrix for transforming pixel offsets: + // video_uv = relative position in video frame, mapped to [0.0, 1.0] range + // tex_uv = relative position in padded texture, mapped to [0.0, 1.0] range + const float2 flat_video_uv = tex_uv * (texture_size * video_size_inv); + float2x2 pixel_to_video_uv; + float2 video_uv_no_geom_overscan; + if(geom_mode > 0.5) + { + video_uv_no_geom_overscan = + get_curved_video_uv_coords_and_tangent_matrix(flat_video_uv, + eye_pos_local, output_size_inv, geom_aspect, + geom_mode, global_to_local, pixel_to_video_uv); + } + else + { + video_uv_no_geom_overscan = flat_video_uv; + pixel_to_video_uv = float2x2( + output_size_inv.x, 0.0, 0.0, output_size_inv.y); + } + // Correct for overscan here (not in curvature code): + const float2 video_uv = + (video_uv_no_geom_overscan - float2(0.5, 0.5))/geom_overscan + float2(0.5, 0.5); + const float2 tex_uv = video_uv * (video_size * texture_size_inv); + + // Get a matrix transforming pixel vectors to tex_uv vectors: + const float2x2 pixel_to_tex_uv = + mul_scale(video_size * texture_size_inv / + geom_aspect_and_overscan.zw, pixel_to_video_uv); + + // Sample! Skip antialiasing if aa_level < 0.5 or both of these hold: + // 1.) Geometry/curvature isn't used + // 2.) Overscan == float2(1.0, 1.0) + // Skipping AA is sharper, but it's only faster with dynamic branches. + const float2 abs_aa_r_offset = abs(get_aa_subpixel_r_offset()); + const bool need_subpixel_aa = abs_aa_r_offset.x + abs_aa_r_offset.y > 0.0; + float3 color; + if(aa_level > 0.5 && (geom_mode > 0.5 || any(bool2((geom_overscan.x != 1.0), (geom_overscan.y != 1.0))))) + { + // Sample the input with antialiasing (due to sharp phosphors, etc.): + color = tex2Daa(input_texture, tex_uv, pixel_to_tex_uv, float(frame_count)); + } + + else if(aa_level > 0.5 && need_subpixel_aa) + { + // Sample at each subpixel location: + color = tex2Daa_subpixel_weights_only( + input_texture, tex_uv, pixel_to_tex_uv); + } + else + { + color = tex2D_linearize(input_texture, tex_uv).rgb; + } + + // Dim borders and output the final result: + const float border_dim_factor = get_border_dim_factor(video_uv, geom_aspect); + const float3 final_color = color * border_dim_factor; + + FragColor = encode_output(float4(final_color, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/geometry-aa-last-pass.vs b/shaders/CRT-Royale.shader/geometry-aa-last-pass.vs new file mode 100644 index 000000000..1c99650d8 --- /dev/null +++ b/shaders/CRT-Royale.shader/geometry-aa-last-pass.vs @@ -0,0 +1,5263 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 tex_uv; + vec4 video_and_texture_size_inv; + vec2 output_size_inv; + vec3 eye_pos_local; + vec4 geom_aspect_and_overscan; + vec3 global_to_local_row0; + vec3 global_to_local_row1; + vec3 global_to_local_row2; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(x,y) +#define rsqrt(c) inversesqrt(c) + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +#define LAST_PASS +#define SIMULATE_CRT_ON_LCD + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +#ifndef RUNTIME_GEOMETRY_TILT + // Create a local-to-global rotation matrix for the CRT's coordinate frame + // and its global-to-local inverse. See the vertex shader for details. + // It's faster to compute these statically if possible. + static const float2 sin_tilt = sin(geom_tilt_angle_static); + static const float2 cos_tilt = cos(geom_tilt_angle_static); + static const float3x3 geom_local_to_global_static = float3x3( + cos_tilt.x, sin_tilt.y*sin_tilt.x, cos_tilt.y*sin_tilt.x, + 0.0, cos_tilt.y, -sin_tilt.y, + -sin_tilt.x, sin_tilt.y*cos_tilt.x, cos_tilt.y*cos_tilt.x); + static const float3x3 geom_global_to_local_static = float3x3( + cos_tilt.x, 0.0, -sin_tilt.x, + sin_tilt.y*sin_tilt.x, cos_tilt.y, sin_tilt.y*cos_tilt.x, + cos_tilt.y*sin_tilt.x, -sin_tilt.y, cos_tilt.y*cos_tilt.x); +#endif + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "tex2Dantialias.h" + +///////////////////////// BEGIN TEX2DANTIALIAS ///////////////////////// + +#ifndef TEX2DANTIALIAS_H +#define TEX2DANTIALIAS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides antialiased and subpixel-aware tex2D lookups. +// Requires: All functions share these requirements: +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) pixel_to_tex_uv must be a 2x2 matrix that transforms pixe- +// space offsets to texture uv offsets. You can get this with: +// const float2 duv_dx = ddx(tex_uv); +// const float2 duv_dy = ddy(tex_uv); +// const float2x2 pixel_to_tex_uv = float2x2( +// duv_dx.x, duv_dy.x, +// duv_dx.y, duv_dy.y); +// This is left to the user in case the current Cg profile +// doesn't support ddx()/ddy(). Ideally, the user could find +// calculate a distorted tangent-space mapping analytically. +// If not, a simple flat mapping can be obtained with: +// const float2 xy_to_uv_scale = output_size * +// video_size/texture_size; +// const float2x2 pixel_to_tex_uv = float2x2( +// xy_to_uv_scale.x, 0.0, +// 0.0, xy_to_uv_scale.y); +// Optional: To set basic AA settings, #define ANTIALIAS_OVERRIDE_BASICS and: +// 1.) Set an antialiasing level: +// static const float aa_level = {0 (none), +// 1 (sample subpixels), 4, 5, 6, 7, 8, 12, 16, 20, 24} +// 2.) Set a filter type: +// static const float aa_filter = { +// 0 (Box, Separable), 1 (Box, Cylindrical), +// 2 (Tent, Separable), 3 (Tent, Cylindrical) +// 4 (Gaussian, Separable), 5 (Gaussian, Cylindrical) +// 6 (Cubic, Separable), 7 (Cubic, Cylindrical) +// 8 (Lanczos Sinc, Separable), +// 9 (Lanczos Jinc, Cylindrical)} +// If the input is unknown, a separable box filter is used. +// Note: Lanczos Jinc is terrible for sparse sampling, and +// using aa_axis_importance (see below) defeats the purpose. +// 3.) Mirror the sample pattern on odd frames? +// static const bool aa_temporal = {true, false] +// This helps rotational invariance but can look "fluttery." +// The user may #define ANTIALIAS_OVERRIDE_PARAMETERS to override +// (all of) the following default parameters with static or uniform +// constants (or an accessor function for subpixel offsets): +// 1.) Cubic parameters: +// static const float aa_cubic_c = 0.5; +// See http://www.imagemagick.org/Usage/filter/#mitchell +// 2.) Gaussian parameters: +// static const float aa_gauss_sigma = +// 0.5/aa_pixel_diameter; +// 3.) Set subpixel offsets. This requires an accessor function +// for compatibility with scalar runtime shader Return +// a float2 pixel offset in [-0.5, 0.5] for the red subpixel: +// float2 get_aa_subpixel_r_offset() +// The user may also #define ANTIALIAS_OVERRIDE_STATIC_CONSTANTS to +// override (all of) the following default static values. However, +// the file's structure requires them to be declared static const: +// 1.) static const float aa_lanczos_lobes = 3.0; +// 2.) static const float aa_gauss_support = 1.0/aa_pixel_diameter; +// Note the default tent/Gaussian support radii may appear +// arbitrary, but extensive testing found them nearly optimal +// for tough cases like strong distortion at low AA levels. +// (The Gaussian default is only best for practical gauss_sigma +// values; much larger gauss_sigmas ironically prefer slightly +// smaller support given sparse sampling, and vice versa.) +// 3.) static const float aa_tent_support = 1.0 / aa_pixel_diameter; +// 4.) static const float2 aa_xy_axis_importance: +// The sparse N-queens sampling grid interacts poorly with +// negative-lobed 2D filters. However, if aliasing is much +// stronger in one direction (e.g. horizontally with a phosphor +// mask), it can be useful to downplay sample offsets along the +// other axis. The support radius in each direction scales with +// aa_xy_axis_importance down to a minimum of 0.5 (box support), +// after which point only the offsets used for calculating +// weights continue to scale downward. This works as follows: +// If aa_xy_axis_importance = float2(1.0, 1.0/support_radius), +// the vertical support radius will drop to 1.0, and we'll just +// filter vertical offsets with the first filter lobe, while +// horizontal offsets go through the full multi-lobe filter. +// If aa_xy_axis_importance = float2(1.0, 0.0), the vertical +// support radius will drop to box support, and the vertical +// offsets will be ignored entirely (essentially giving us a +// box filter vertically). The former is potentially smoother +// (but less predictable) and the default behavior of Lanczos +// jinc, whereas the latter is sharper and the default behavior +// of cubics and Lanczos sinc. +// 5.) static const float aa_pixel_diameter: You can expand the +// pixel diameter to e.g. sqrt(2.0), which may be a better +// support range for cylindrical filters (they don't +// currently discard out-of-circle samples though). +// Finally, there are two miscellaneous options: +// 1.) If you want to antialias a manually tiled texture, you can +// #define ANTIALIAS_DISABLE_ANISOTROPIC to use tex2Dlod() to +// fix incompatibilities with anisotropic filtering. This is +// slower, and the Cg profile must support tex2Dlod(). +// 2.) If aa_cubic_c is a runtime uniform, you can #define +// RUNTIME_ANTIALIAS_WEIGHTS to evaluate cubic weights once per +// fragment instead of at the usage site (which is used by +// default, because it enables static evaluation). +// Description: +// Each antialiased lookup follows these steps: +// 1.) Define a sample pattern of pixel offsets in the range of [-0.5, 0.5] +// pixels, spanning the diameter of a rectangular box filter. +// 2.) Scale these offsets by the support diameter of the user's chosen filter. +// 3.) Using these pixel offsets from the pixel center, compute the offsets to +// predefined subpixel locations. +// 4.) Compute filter weights based on subpixel offsets. +// Much of that can often be done at compile-time. At runtime: +// 1.) Project pixel-space offsets into uv-space with a matrix multiplication +// to get the uv offsets for each sample. Rectangular pixels have a +// diameter of 1.0. Circular pixels are not currently supported, but they +// might be better with a diameter of sqrt(2.0) to ensure there are no gaps +// between them. +// 2.) Load, weight, and sum samples. +// We use a sparse bilinear sampling grid, so there are two major implications: +// 1.) We can directly project the pixel-space support box into uv-space even +// if we're upsizing. This wouldn't be the case for nearest neighbor, +// where we'd have to expand the uv-space diameter to at least the support +// size to ensure sufficient filter support. In our case, this allows us +// to treat upsizing the same as downsizing and use static weighting. :) +// 2.) For decent results, negative-lobed filters must be computed based on +// separable weights, not radial distances, because the sparse sampling +// makes no guarantees about radial distributions. Even then, it's much +// better to set aa_xy_axis_importance to e.g. float2(1.0, 0.0) to use e.g. +// Lanczos2 horizontally and a box filter vertically. This is mainly due +// to the sparse N-queens sampling and a statistically enormous positive or +// negative covariance between horizontal and vertical weights. +// +// Design Decision Comments: +// "aa_temporal" mirrors the sample pattern on odd frames along the axis that +// keeps subpixel weights constant. This helps with rotational invariance, but +// it can cause distracting fluctuations, and horizontal and vertical edges +// will look the same. Using a different pattern on a shifted grid would +// exploit temporal AA better, but it would require a dynamic branch or a lot +// of conditional moves, so it's prohibitively slow for the minor benefit. + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +#ifndef ANTIALIAS_OVERRIDE_BASICS + // The following settings must be static constants: + static const float aa_level = 12.0; + static const float aa_filter = 0.0; + static const bool aa_temporal = false; +#endif + +#ifndef ANTIALIAS_OVERRIDE_STATIC_CONSTANTS + // Users may override these parameters, but the file structure requires + // them to be static constants; see the descriptions above. + static const float aa_pixel_diameter = 1.0; + static const float aa_lanczos_lobes = 3.0; + static const float aa_gauss_support = 1.0 / aa_pixel_diameter; + static const float aa_tent_support = 1.0 / aa_pixel_diameter; + + // If we're using a negative-lobed filter, default to using it horizontally + // only, and use only the first lobe vertically or a box filter, over a + // correspondingly smaller range. This compensates for the sparse sampling + // grid's typically large positive/negative x/y covariance. + static const float2 aa_xy_axis_importance = + aa_filter < 5.5 ? float2(1.0) : // Box, tent, Gaussian + aa_filter < 8.5 ? float2(1.0, 0.0) : // Cubic and Lanczos sinc + aa_filter < 9.5 ? float2(1.0, 1.0/aa_lanczos_lobes) : // Lanczos jinc + float2(1.0); // Default to box +#endif + +#ifndef ANTIALIAS_OVERRIDE_PARAMETERS + // Users may override these values with their own uniform or static consts. + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c = 0.5; + static const float aa_gauss_sigma = 0.5 / aa_pixel_diameter; + // Users may override the subpixel offset accessor function with their own. + // A function is used for compatibility with scalar runtime shader + inline float2 get_aa_subpixel_r_offset() + { + return float2(0.0, 0.0); + } +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../../../../include/gamma-management.h" + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +static const float aa_box_support = 0.5; +static const float aa_cubic_support = 2.0; + + +//////////////////////////// GLOBAL NON-CONSTANTS //////////////////////////// + +// We'll want to define these only once per fragment at most. +#ifdef RUNTIME_ANTIALIAS_WEIGHTS + float aa_cubic_b; + float cubic_branch1_x3_coeff; + float cubic_branch1_x2_coeff; + float cubic_branch1_x0_coeff; + float cubic_branch2_x3_coeff; + float cubic_branch2_x2_coeff; + float cubic_branch2_x1_coeff; + float cubic_branch2_x0_coeff; +#endif + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +void assign_aa_cubic_constants() +{ + // Compute cubic coefficients on demand at runtime, and save them to global + // uniforms. The B parameter is computed from C, because "Keys cubics" + // with B = 1 - 2C are considered the highest quality. + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + if(aa_filter > 5.5 && aa_filter < 7.5) + { + aa_cubic_b = 1.0 - 2.0*aa_cubic_c; + cubic_branch1_x3_coeff = 12.0 - 9.0*aa_cubic_b - 6.0*aa_cubic_c; + cubic_branch1_x2_coeff = -18.0 + 12.0*aa_cubic_b + 6.0*aa_cubic_c; + cubic_branch1_x0_coeff = 6.0 - 2.0 * aa_cubic_b; + cubic_branch2_x3_coeff = -aa_cubic_b - 6.0 * aa_cubic_c; + cubic_branch2_x2_coeff = 6.0*aa_cubic_b + 30.0*aa_cubic_c; + cubic_branch2_x1_coeff = -12.0*aa_cubic_b - 48.0*aa_cubic_c; + cubic_branch2_x0_coeff = 8.0*aa_cubic_b + 24.0*aa_cubic_c; + } + #endif +} + +inline float4 get_subpixel_support_diam_and_final_axis_importance() +{ + // Statically select the base support radius: + static const float base_support_radius = + aa_filter < 1.5 ? aa_box_support : + aa_filter < 3.5 ? aa_tent_support : + aa_filter < 5.5 ? aa_gauss_support : + aa_filter < 7.5 ? aa_cubic_support : + aa_filter < 9.5 ? aa_lanczos_lobes : + aa_box_support; // Default to box + // Expand the filter support for subpixel filtering. + const float2 subpixel_support_radius_raw = + float2(base_support_radius) + abs(get_aa_subpixel_r_offset()); + if(aa_filter < 1.5) + { + // Ignore aa_xy_axis_importance for box filtering. + const float2 subpixel_support_diam = + 2.0 * subpixel_support_radius_raw; + const float2 final_axis_importance = float2(1.0); + return float4(subpixel_support_diam, final_axis_importance); + } + else + { + // Scale the support window by aa_xy_axis_importance, but don't narrow + // it further than box support. This allows decent vertical AA without + // messing up horizontal weights or using something silly like Lanczos4 + // horizontally with a huge vertical average over an 8-pixel radius. + const float2 subpixel_support_radius = max(float2(aa_box_support, aa_box_support), + subpixel_support_radius_raw * aa_xy_axis_importance); + // Adjust aa_xy_axis_importance to compensate for what's already done: + const float2 final_axis_importance = aa_xy_axis_importance * + subpixel_support_radius_raw/subpixel_support_radius; + const float2 subpixel_support_diam = 2.0 * subpixel_support_radius; + return float4(subpixel_support_diam, final_axis_importance); + } +} + + +/////////////////////////// FILTER WEIGHT FUNCTIONS ////////////////////////// + +inline float eval_box_filter(const float dist) +{ + return float(abs(dist) <= aa_box_support); +} + +inline float eval_separable_box_filter(const float2 offset) +{ + return float(all(bool2((abs(offset.x) <= aa_box_support), (abs(offset.y) <= aa_box_support)))); +} + +inline float eval_tent_filter(const float dist) +{ + return clamp((aa_tent_support - dist)/ + aa_tent_support, 0.0, 1.0); +} + +inline float eval_gaussian_filter(const float dist) +{ + return exp(-(dist*dist) / (2.0*aa_gauss_sigma*aa_gauss_sigma)); +} + +inline float eval_cubic_filter(const float dist) +{ + // Compute coefficients like assign_aa_cubic_constants(), but statically. + #ifndef RUNTIME_ANTIALIAS_WEIGHTS + // When runtime weights are used, these values are instead written to + // global uniforms at the beginning of each tex2Daa* call. + const float aa_cubic_b = 1.0 - 2.0*aa_cubic_c; + const float cubic_branch1_x3_coeff = 12.0 - 9.0*aa_cubic_b - 6.0*aa_cubic_c; + const float cubic_branch1_x2_coeff = -18.0 + 12.0*aa_cubic_b + 6.0*aa_cubic_c; + const float cubic_branch1_x0_coeff = 6.0 - 2.0 * aa_cubic_b; + const float cubic_branch2_x3_coeff = -aa_cubic_b - 6.0 * aa_cubic_c; + const float cubic_branch2_x2_coeff = 6.0*aa_cubic_b + 30.0*aa_cubic_c; + const float cubic_branch2_x1_coeff = -12.0*aa_cubic_b - 48.0*aa_cubic_c; + const float cubic_branch2_x0_coeff = 8.0*aa_cubic_b + 24.0*aa_cubic_c; + #endif + const float abs_dist = abs(dist); + // Compute the cubic based on the Horner's method formula in: + // http://www.cs.utexas.edu/users/fussell/courses/cs384g/lectures/mitchell/Mitchell.pdf + return (abs_dist < 1.0 ? + (cubic_branch1_x3_coeff*abs_dist + + cubic_branch1_x2_coeff)*abs_dist*abs_dist + + cubic_branch1_x0_coeff : + abs_dist < 2.0 ? + ((cubic_branch2_x3_coeff*abs_dist + + cubic_branch2_x2_coeff)*abs_dist + + cubic_branch2_x1_coeff)*abs_dist + cubic_branch2_x0_coeff : + 0.0)/6.0; +} + +inline float eval_separable_cubic_filter(const float2 offset) +{ + // This is faster than using a specific float2 version: + return eval_cubic_filter(offset.x) * + eval_cubic_filter(offset.y); +} + +inline float2 eval_sinc_filter(const float2 offset) +{ + // It's faster to let the caller handle the zero case, or at least it + // was when I used macros and the shader preset took a full minute to load. + const float2 pi_offset = pi * offset; + return sin(pi_offset)/pi_offset; +} + +inline float eval_separable_lanczos_sinc_filter(const float2 offset_unsafe) +{ + // Note: For sparse sampling, you really need to pick an axis to use + // Lanczos along (e.g. set aa_xy_axis_importance = float2(1.0, 0.0)). + const float2 offset = FIX_ZERO(offset_unsafe); + const float2 xy_weights = eval_sinc_filter(offset) * + eval_sinc_filter(offset/aa_lanczos_lobes); + return xy_weights.x * xy_weights.y; +} + +inline float eval_jinc_filter_unorm(const float x) +{ + // This is a Jinc approximation for x in [0, 45). We'll use x in range + // [0, 4*pi) or so. There are faster/closer approximations based on + // piecewise cubics from [0, 45) and asymptotic approximations beyond that, + // but this has a maximum absolute error < 1/512, and it's simpler/faster + // for shaders...not that it's all that useful for sparse sampling anyway. + const float point3845_x = 0.38448566093564*x; + const float exp_term = exp(-(point3845_x*point3845_x)); + const float point8154_plus_x = 0.815362332840791 + x; + const float cos_term = cos(point8154_plus_x); + return ( + 0.0264727330997042*min(x, 6.83134964622778) + + 0.680823557250528*exp_term + + -0.0597255978950933*min(7.41043194481873, x)*cos_term / + (point8154_plus_x + 0.0646074538634482*(x*x) + + cos(x)*max(exp_term, cos(x) + cos_term)) - + 0.180837503591406); +} + +inline float eval_jinc_filter(const float dist) +{ + return eval_jinc_filter_unorm(pi * dist); +} + +inline float eval_lanczos_jinc_filter(const float dist) +{ + return eval_jinc_filter(dist) * eval_jinc_filter(dist/aa_lanczos_lobes); +} + + +inline float3 eval_unorm_rgb_weights(const float2 offset, + const float2 final_axis_importance) +{ + // Requires: 1.) final_axis_impportance must be computed according to + // get_subpixel_support_diam_and_final_axis_importance(). + // 2.) aa_filter must be a global constant. + // 3.) offset must be an xy pixel offset in the range: + // ([-subpixel_support_diameter.x/2, + // subpixel_support_diameter.x/2], + // [-subpixel_support_diameter.y/2, + // subpixel_support_diameter.y/2]) + // Returns: Sample weights at R/G/B destination subpixels for the + // given xy pixel offset. + const float2 offset_g = offset * final_axis_importance; + const float2 aa_r_offset = get_aa_subpixel_r_offset(); + const float2 offset_r = offset_g - aa_r_offset * final_axis_importance; + const float2 offset_b = offset_g + aa_r_offset * final_axis_importance; + // Statically select a filter: + if(aa_filter < 0.5) + { + return float3(eval_separable_box_filter(offset_r), + eval_separable_box_filter(offset_g), + eval_separable_box_filter(offset_b)); + } + else if(aa_filter < 1.5) + { + return float3(eval_box_filter(length(offset_r)), + eval_box_filter(length(offset_g)), + eval_box_filter(length(offset_b))); + } + else if(aa_filter < 2.5) + { + return float3( + eval_tent_filter(offset_r.x) * eval_tent_filter(offset_r.y), + eval_tent_filter(offset_g.x) * eval_tent_filter(offset_g.y), + eval_tent_filter(offset_b.x) * eval_tent_filter(offset_b.y)); + } + else if(aa_filter < 3.5) + { + return float3(eval_tent_filter(length(offset_r)), + eval_tent_filter(length(offset_g)), + eval_tent_filter(length(offset_b))); + } + else if(aa_filter < 4.5) + { + return float3( + eval_gaussian_filter(offset_r.x) * eval_gaussian_filter(offset_r.y), + eval_gaussian_filter(offset_g.x) * eval_gaussian_filter(offset_g.y), + eval_gaussian_filter(offset_b.x) * eval_gaussian_filter(offset_b.y)); + } + else if(aa_filter < 5.5) + { + return float3(eval_gaussian_filter(length(offset_r)), + eval_gaussian_filter(length(offset_g)), + eval_gaussian_filter(length(offset_b))); + } + else if(aa_filter < 6.5) + { + return float3( + eval_cubic_filter(offset_r.x) * eval_cubic_filter(offset_r.y), + eval_cubic_filter(offset_g.x) * eval_cubic_filter(offset_g.y), + eval_cubic_filter(offset_b.x) * eval_cubic_filter(offset_b.y)); + } + else if(aa_filter < 7.5) + { + return float3(eval_cubic_filter(length(offset_r)), + eval_cubic_filter(length(offset_g)), + eval_cubic_filter(length(offset_b))); + } + else if(aa_filter < 8.5) + { + return float3(eval_separable_lanczos_sinc_filter(offset_r), + eval_separable_lanczos_sinc_filter(offset_g), + eval_separable_lanczos_sinc_filter(offset_b)); + } + else if(aa_filter < 9.5) + { + return float3(eval_lanczos_jinc_filter(length(offset_r)), + eval_lanczos_jinc_filter(length(offset_g)), + eval_lanczos_jinc_filter(length(offset_b))); + } + else + { + // Default to a box, because Lanczos Jinc is so bad. ;) + return float3(eval_separable_box_filter(offset_r), + eval_separable_box_filter(offset_g), + eval_separable_box_filter(offset_b)); + } +} + + +////////////////////////////// HELPER FUNCTIONS ////////////////////////////// + +inline float4 tex2Daa_tiled_linearize(const sampler2D samp, const float2 s) +{ + // If we're manually tiling a texture, anisotropic filtering can get + // confused. This is one workaround: + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + // TODO: Use tex2Dlod_linearize with a calculated mip level. + return tex2Dlod_linearize(samp, float4(s, 0.0, 0.0)); + #else + return tex2D_linearize(samp, s); + #endif +} + +inline float2 get_frame_sign(const float frame) +{ + if(aa_temporal) + { + // Mirror the sampling pattern for odd frames in a direction that + // lets us keep the same subpixel sample weights: + const float frame_odd = float(fmod(frame, 2.0) > 0.5); + const float2 aa_r_offset = get_aa_subpixel_r_offset(); + const float2 mirror = -float2(abs(aa_r_offset.x) < (FIX_ZERO(0.0)), abs(aa_r_offset.y) < (FIX_ZERO(0.0))); + return mirror; + } + else + { + return float2(1.0, 1.0); + } +} + + +///////////////////////// ANTIALIASED TEXTURE LOOKUPS //////////////////////// + +float3 tex2Daa_subpixel_weights_only(const sampler2D tex, + const float2 tex_uv, const float2x2 pixel_to_tex_uv) +{ + // This function is unlike the others: Just perform a single independent + // lookup for each subpixel. It may be very aliased. + const float2 aa_r_offset = get_aa_subpixel_r_offset(); + const float2 aa_r_offset_uv_offset = mul(pixel_to_tex_uv, aa_r_offset); + const float color_g = tex2D_linearize(tex, tex_uv).g; + const float color_r = tex2D_linearize(tex, tex_uv + aa_r_offset_uv_offset).r; + const float color_b = tex2D_linearize(tex, tex_uv - aa_r_offset_uv_offset).b; + return float3(color_r, color_g, color_b); +} + +// The tex2Daa* functions compile very slowly due to all the macros and +// compile-time math, so only include the ones we'll actually use! +float3 tex2Daa4x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use an RGMS4 pattern (4-queens): + // . . Q . : off =(-1.5, -1.5)/4 + (2.0, 0.0)/4 + // Q . . . : off =(-1.5, -1.5)/4 + (0.0, 1.0)/4 + // . . . Q : off =(-1.5, -1.5)/4 + (3.0, 2.0)/4 + // . Q . . : off =(-1.5, -1.5)/4 + (1.0, 3.0)/4 + // Static screenspace sample offsets (compute some implicitly): + static const float grid_size = 4.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0,1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5,0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(0.0, 1.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = w1.bgr; + const float3 w3 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0,1.0,1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * (w0 * sample0 + w1 * sample1 + + w2 * sample2 + w3 * sample3); +} + +float3 tex2Daa5x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 5-queens pattern: + // . Q . . . : off =(-2.0, -2.0)/5 + (1.0, 0.0)/5 + // . . . . Q : off =(-2.0, -2.0)/5 + (4.0, 1.0)/5 + // . . Q . . : off =(-2.0, -2.0)/5 + (2.0, 2.0)/5 + // Q . . . . : off =(-2.0, -2.0)/5 + (0.0, 3.0)/5 + // . . . Q . : off =(-2.0, -2.0)/5 + (3.0, 4.0)/5 + // Static screenspace sample offsets (compute some implicitly): + static const float grid_size = 5.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(1.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(4.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(2.0, 2.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = w1.bgr; + const float3 w4 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 w_sum_inv = float3(1.0)/(w0 + w1 + w2 + w3 + w4); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * (w0 * sample0 + w1 * sample1 + + w2 * sample2 + w3 * sample3 + w4 * sample4); +} + +float3 tex2Daa6x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 6-queens pattern with a stronger horizontal + // than vertical slant: + // . . . . Q . : off =(-2.5, -2.5)/6 + (4.0, 0.0)/6 + // . . Q . . . : off =(-2.5, -2.5)/6 + (2.0, 1.0)/6 + // Q . . . . . : off =(-2.5, -2.5)/6 + (0.0, 2.0)/6 + // . . . . . Q : off =(-2.5, -2.5)/6 + (5.0, 3.0)/6 + // . . . Q . . : off =(-2.5, -2.5)/6 + (3.0, 4.0)/6 + // . Q . . . . : off =(-2.5, -2.5)/6 + (1.0, 5.0)/6 + // Static screenspace sample offsets (compute some implicitly): + static const float grid_size = 6.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(4.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(2.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(0.0, 2.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = w2.bgr; + const float3 w4 = w1.bgr; + const float3 w5 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * (w0 * sample0 + w1 * sample1 + w2 * sample2 + + w3 * sample3 + w4 * sample4 + w5 * sample5); +} + +float3 tex2Daa7x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 7-queens pattern with a queen in the center: + // . Q . . . . . : off =(-3.0, -3.0)/7 + (1.0, 0.0)/7 + // . . . . Q . . : off =(-3.0, -3.0)/7 + (4.0, 1.0)/7 + // Q . . . . . . : off =(-3.0, -3.0)/7 + (0.0, 2.0)/7 + // . . . Q . . . : off =(-3.0, -3.0)/7 + (3.0, 3.0)/7 + // . . . . . . Q : off =(-3.0, -3.0)/7 + (6.0, 4.0)/7 + // . . Q . . . . : off =(-3.0, -3.0)/7 + (2.0, 5.0)/7 + // . . . . . Q . : off =(-3.0, -3.0)/7 + (5.0, 6.0)/7 + static const float grid_size = 7.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(1.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(4.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(0.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(3.0, 3.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = w2.bgr; + const float3 w5 = w1.bgr; + const float3 w6 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2; + const float3 w_sum = half_sum + half_sum.bgr + w3; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6); +} + +float3 tex2Daa8x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 8-queens pattern. + // . . Q . . . . . : off =(-3.5, -3.5)/8 + (2.0, 0.0)/8 + // . . . . Q . . . : off =(-3.5, -3.5)/8 + (4.0, 1.0)/8 + // . Q . . . . . . : off =(-3.5, -3.5)/8 + (1.0, 2.0)/8 + // . . . . . . . Q : off =(-3.5, -3.5)/8 + (7.0, 3.0)/8 + // Q . . . . . . . : off =(-3.5, -3.5)/8 + (0.0, 4.0)/8 + // . . . . . . Q . : off =(-3.5, -3.5)/8 + (6.0, 5.0)/8 + // . . . Q . . . . : off =(-3.5, -3.5)/8 + (3.0, 6.0)/8 + // . . . . . Q . . : off =(-3.5, -3.5)/8 + (5.0, 7.0)/8 + static const float grid_size = 8.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(4.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(1.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(7.0, 3.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = w3.bgr; + const float3 w5 = w2.bgr; + const float3 w6 = w1.bgr; + const float3 w7 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, and mirror on odd frames if directed: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7); +} + +float3 tex2Daa12x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 12-superqueens pattern where no 3 points are + // exactly collinear. + // . . . Q . . . . . . . . : off =(-5.5, -5.5)/12 + (3.0, 0.0)/12 + // . . . . . . . . . Q . . : off =(-5.5, -5.5)/12 + (9.0, 1.0)/12 + // . . . . . . Q . . . . . : off =(-5.5, -5.5)/12 + (6.0, 2.0)/12 + // . Q . . . . . . . . . . : off =(-5.5, -5.5)/12 + (1.0, 3.0)/12 + // . . . . . . . . . . . Q : off =(-5.5, -5.5)/12 + (11.0, 4.0)/12 + // . . . . Q . . . . . . . : off =(-5.5, -5.5)/12 + (4.0, 5.0)/12 + // . . . . . . . Q . . . . : off =(-5.5, -5.5)/12 + (7.0, 6.0)/12 + // Q . . . . . . . . . . . : off =(-5.5, -5.5)/12 + (0.0, 7.0)/12 + // . . . . . . . . . . Q . : off =(-5.5, -5.5)/12 + (10.0, 8.0)/12 + // . . . . . Q . . . . . . : off =(-5.5, -5.5)/12 + (5.0, 9.0)/12 + // . . Q . . . . . . . . . : off =(-5.5, -5.5)/12 + (2.0, 10.0)/12 + // . . . . . . . . Q . . . : off =(-5.5, -5.5)/12 + (8.0, 11.0)/12 + static const float grid_size = 12.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(3.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(9.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(6.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(1.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(11.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(4.0, 5.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = w5.bgr; + const float3 w7 = w4.bgr; + const float3 w8 = w3.bgr; + const float3 w9 = w2.bgr; + const float3 w10 = w1.bgr; + const float3 w11 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/w_sum; + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11); +} + +float3 tex2Daa16x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 16-superqueens pattern where no 3 points are + // exactly collinear. + // . . Q . . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (2.0, 0.0)/16 + // . . . . . . . . . Q . . . . . . : off =(-7.5, -7.5)/16 + (9.0, 1.0)/16 + // . . . . . . . . . . . . Q . . . : off =(-7.5, -7.5)/16 + (12.0, 2.0)/16 + // . . . . Q . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (4.0, 3.0)/16 + // . . . . . . . . Q . . . . . . . : off =(-7.5, -7.5)/16 + (8.0, 4.0)/16 + // . . . . . . . . . . . . . . Q . : off =(-7.5, -7.5)/16 + (14.0, 5.0)/16 + // Q . . . . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (0.0, 6.0)/16 + // . . . . . . . . . . Q . . . . . : off =(-7.5, -7.5)/16 + (10.0, 7.0)/16 + // . . . . . Q . . . . . . . . . . : off =(-7.5, -7.5)/16 + (5.0, 8.0)/16 + // . . . . . . . . . . . . . . . Q : off =(-7.5, -7.5)/16 + (15.0, 9.0)/16 + // . Q . . . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (1.0, 10.0)/16 + // . . . . . . . Q . . . . . . . . : off =(-7.5, -7.5)/16 + (7.0, 11.0)/16 + // . . . . . . . . . . . Q . . . . : off =(-7.5, -7.5)/16 + (11.0, 12.0)/16 + // . . . Q . . . . . . . . . . . . : off =(-7.5, -7.5)/16 + (3.0, 13.0)/16 + // . . . . . . Q . . . . . . . . . : off =(-7.5, -7.5)/16 + (6.0, 14.0)/16 + // . . . . . . . . . . . . . Q . . : off =(-7.5, -7.5)/16 + (13.0, 15.0)/16 + static const float grid_size = 16.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(9.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(12.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(4.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(8.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(14.0, 5.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(0.0, 6.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(10.0, 7.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = w7.bgr; + const float3 w9 = w6.bgr; + const float3 w10 = w5.bgr; + const float3 w11 = w4.bgr; + const float3 w12 = w3.bgr; + const float3 w13 = w2.bgr; + const float3 w14 = w1.bgr; + const float3 w15 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + const float2 uv_offset6 = mul(true_pixel_to_tex_uv, xy_offset6 * frame_sign); + const float2 uv_offset7 = mul(true_pixel_to_tex_uv, xy_offset7 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset6).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset7).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset7).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset6).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15); +} + +float3 tex2Daa20x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 20-superqueens pattern where no 3 points are + // exactly collinear and superqueens have a squared attack radius of 13. + // . . . . . . . Q . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (7.0, 0.0)/20 + // . . . . . . . . . . . . . . . . Q . . . : off =(-9.5, -9.5)/20 + (16.0, 1.0)/20 + // . . . . . . . . . . . Q . . . . . . . . : off =(-9.5, -9.5)/20 + (11.0, 2.0)/20 + // . Q . . . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (1.0, 3.0)/20 + // . . . . . Q . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (5.0, 4.0)/20 + // . . . . . . . . . . . . . . . Q . . . . : off =(-9.5, -9.5)/20 + (15.0, 5.0)/20 + // . . . . . . . . . . Q . . . . . . . . . : off =(-9.5, -9.5)/20 + (10.0, 6.0)/20 + // . . . . . . . . . . . . . . . . . . . Q : off =(-9.5, -9.5)/20 + (19.0, 7.0)/20 + // . . Q . . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (2.0, 8.0)/20 + // . . . . . . Q . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (6.0, 9.0)/20 + // . . . . . . . . . . . . . Q . . . . . . : off =(-9.5, -9.5)/20 + (13.0, 10.0)/20 + // . . . . . . . . . . . . . . . . . Q . . : off =(-9.5, -9.5)/20 + (17.0, 11.0)/20 + // Q . . . . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (0.0, 12.0)/20 + // . . . . . . . . . Q . . . . . . . . . . : off =(-9.5, -9.5)/20 + (9.0, 13.0)/20 + // . . . . Q . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (4.0, 14.0)/20 + // . . . . . . . . . . . . . . Q . . . . . : off =(-9.5, -9.5)/20 + (14.0, 15.0)/20 + // . . . . . . . . . . . . . . . . . . Q . : off =(-9.5, -9.5)/20 + (18.0, 16.0)/20 + // . . . . . . . . Q . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (8.0, 17.0)/20 + // . . . Q . . . . . . . . . . . . . . . . : off =(-9.5, -9.5)/20 + (3.0, 18.0)/20 + // . . . . . . . . . . . . Q . . . . . . . : off =(-9.5, -9.5)/20 + (12.0, 19.0)/20 + static const float grid_size = 20.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(7.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(16.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(11.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(1.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(5.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(15.0, 5.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(10.0, 6.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(19.0, 7.0) * xy_step; + const float2 xy_offset8 = xy_start_offset + float2(2.0, 8.0) * xy_step; + const float2 xy_offset9 = xy_start_offset + float2(6.0, 9.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = eval_unorm_rgb_weights(xy_offset8, final_axis_importance); + const float3 w9 = eval_unorm_rgb_weights(xy_offset9, final_axis_importance); + const float3 w10 = w9.bgr; + const float3 w11 = w8.bgr; + const float3 w12 = w7.bgr; + const float3 w13 = w6.bgr; + const float3 w14 = w5.bgr; + const float3 w15 = w4.bgr; + const float3 w16 = w3.bgr; + const float3 w17 = w2.bgr; + const float3 w18 = w1.bgr; + const float3 w19 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + const float2 uv_offset6 = mul(true_pixel_to_tex_uv, xy_offset6 * frame_sign); + const float2 uv_offset7 = mul(true_pixel_to_tex_uv, xy_offset7 * frame_sign); + const float2 uv_offset8 = mul(true_pixel_to_tex_uv, xy_offset8 * frame_sign); + const float2 uv_offset9 = mul(true_pixel_to_tex_uv, xy_offset9 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset6).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset7).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset8).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset9).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset9).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset8).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset7).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset6).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample16 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample17 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample18 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample19 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15 + + w16 * sample16 + w17 * sample17 + w18 * sample18 + w19 * sample19); +} + +float3 tex2Daa24x(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Use a diagonally symmetric 24-superqueens pattern where no 3 points are + // exactly collinear and superqueens have a squared attack radius of 13. + // . . . . . . Q . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (6.0, 0.0)/24 + // . . . . . . . . . . . . . . . . Q . . . . . . . : off =(-11.5, -11.5)/24 + (16.0, 1.0)/24 + // . . . . . . . . . . Q . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (10.0, 2.0)/24 + // . . . . . . . . . . . . . . . . . . . . . Q . . : off =(-11.5, -11.5)/24 + (21.0, 3.0)/24 + // . . . . . Q . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (5.0, 4.0)/24 + // . . . . . . . . . . . . . . . Q . . . . . . . . : off =(-11.5, -11.5)/24 + (15.0, 5.0)/24 + // . Q . . . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (1.0, 6.0)/24 + // . . . . . . . . . . . Q . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (11.0, 7.0)/24 + // . . . . . . . . . . . . . . . . . . . Q . . . . : off =(-11.5, -11.5)/24 + (19.0, 8.0)/24 + // . . . . . . . . . . . . . . . . . . . . . . . Q : off =(-11.5, -11.5)/24 + (23.0, 9.0)/24 + // . . . Q . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (3.0, 10.0)/24 + // . . . . . . . . . . . . . . Q . . . . . . . . . : off =(-11.5, -11.5)/24 + (14.0, 11.0)/24 + // . . . . . . . . . Q . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (9.0, 12.0)/24 + // . . . . . . . . . . . . . . . . . . . . Q . . . : off =(-11.5, -11.5)/24 + (20.0, 13.0)/24 + // Q . . . . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (0.0, 14.0)/24 + // . . . . Q . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (4.0, 15.0)/24 + // . . . . . . . . . . . . Q . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (12.0, 16.0)/24 + // . . . . . . . . . . . . . . . . . . . . . . Q . : off =(-11.5, -11.5)/24 + (22.0, 17.0)/24 + // . . . . . . . . Q . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (8.0, 18.0)/24 + // . . . . . . . . . . . . . . . . . . Q . . . . . : off =(-11.5, -11.5)/24 + (18.0, 19.0)/24 + // . . Q . . . . . . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (2.0, 20.0)/24 + // . . . . . . . . . . . . . Q . . . . . . . . . . : off =(-11.5, -11.5)/24 + (13.0, 21.0)/24 + // . . . . . . . Q . . . . . . . . . . . . . . . . : off =(-11.5, -11.5)/24 + (7.0, 22.0)/24 + // . . . . . . . . . . . . . . . . . Q . . . . . . : off =(-11.5, -11.5)/24 + (17.0, 23.0)/24 + static const float grid_size = 24.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample. Exploit diagonal symmetry: + const float2 xy_offset0 = xy_start_offset + float2(6.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(16.0, 1.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(10.0, 2.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(21.0, 3.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(5.0, 4.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(15.0, 5.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(1.0, 6.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(11.0, 7.0) * xy_step; + const float2 xy_offset8 = xy_start_offset + float2(19.0, 8.0) * xy_step; + const float2 xy_offset9 = xy_start_offset + float2(23.0, 9.0) * xy_step; + const float2 xy_offset10 = xy_start_offset + float2(3.0, 10.0) * xy_step; + const float2 xy_offset11 = xy_start_offset + float2(14.0, 11.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = eval_unorm_rgb_weights(xy_offset8, final_axis_importance); + const float3 w9 = eval_unorm_rgb_weights(xy_offset9, final_axis_importance); + const float3 w10 = eval_unorm_rgb_weights(xy_offset10, final_axis_importance); + const float3 w11 = eval_unorm_rgb_weights(xy_offset11, final_axis_importance); + const float3 w12 = w11.bgr; + const float3 w13 = w10.bgr; + const float3 w14 = w9.bgr; + const float3 w15 = w8.bgr; + const float3 w16 = w7.bgr; + const float3 w17 = w6.bgr; + const float3 w18 = w5.bgr; + const float3 w19 = w4.bgr; + const float3 w20 = w3.bgr; + const float3 w21 = w2.bgr; + const float3 w22 = w1.bgr; + const float3 w23 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + + w5 + w6 + w7 + w8 + w9 + w10 + w11; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, mirror on odd frames if directed, and exploit + // diagonal symmetry: + const float2 frame_sign = get_frame_sign(frame); + const float2 uv_offset0 = mul(true_pixel_to_tex_uv, xy_offset0 * frame_sign); + const float2 uv_offset1 = mul(true_pixel_to_tex_uv, xy_offset1 * frame_sign); + const float2 uv_offset2 = mul(true_pixel_to_tex_uv, xy_offset2 * frame_sign); + const float2 uv_offset3 = mul(true_pixel_to_tex_uv, xy_offset3 * frame_sign); + const float2 uv_offset4 = mul(true_pixel_to_tex_uv, xy_offset4 * frame_sign); + const float2 uv_offset5 = mul(true_pixel_to_tex_uv, xy_offset5 * frame_sign); + const float2 uv_offset6 = mul(true_pixel_to_tex_uv, xy_offset6 * frame_sign); + const float2 uv_offset7 = mul(true_pixel_to_tex_uv, xy_offset7 * frame_sign); + const float2 uv_offset8 = mul(true_pixel_to_tex_uv, xy_offset8 * frame_sign); + const float2 uv_offset9 = mul(true_pixel_to_tex_uv, xy_offset9 * frame_sign); + const float2 uv_offset10 = mul(true_pixel_to_tex_uv, xy_offset10 * frame_sign); + const float2 uv_offset11 = mul(true_pixel_to_tex_uv, xy_offset11 * frame_sign); + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset0).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset1).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset2).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset3).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset4).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset5).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset6).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset7).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset8).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset9).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset10).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, tex_uv + uv_offset11).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset11).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset10).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset9).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset8).rgb; + const float3 sample16 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset7).rgb; + const float3 sample17 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset6).rgb; + const float3 sample18 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset5).rgb; + const float3 sample19 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset4).rgb; + const float3 sample20 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset3).rgb; + const float3 sample21 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset2).rgb; + const float3 sample22 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset1).rgb; + const float3 sample23 = tex2Daa_tiled_linearize(tex, tex_uv - uv_offset0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15 + + w16 * sample16 + w17 * sample17 + w18 * sample18 + w19 * sample19 + + w20 * sample20 + w21 * sample21 + w22 * sample22 + w23 * sample23); +} + +float3 tex2Daa_debug_16x_regular(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // Sample on a regular 4x4 grid. This is mainly for testing. + static const float grid_size = 4.0; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float2 xy_step = float2(1.0)/grid_size * subpixel_support_diameter; + const float2 xy_start_offset = float2(0.5 - grid_size*0.5) * xy_step; + // Get the xy offset of each sample: + const float2 xy_offset0 = xy_start_offset + float2(0.0, 0.0) * xy_step; + const float2 xy_offset1 = xy_start_offset + float2(1.0, 0.0) * xy_step; + const float2 xy_offset2 = xy_start_offset + float2(2.0, 0.0) * xy_step; + const float2 xy_offset3 = xy_start_offset + float2(3.0, 0.0) * xy_step; + const float2 xy_offset4 = xy_start_offset + float2(0.0, 1.0) * xy_step; + const float2 xy_offset5 = xy_start_offset + float2(1.0, 1.0) * xy_step; + const float2 xy_offset6 = xy_start_offset + float2(2.0, 1.0) * xy_step; + const float2 xy_offset7 = xy_start_offset + float2(3.0, 1.0) * xy_step; + // Compute subpixel weights, and exploit diagonal symmetry for speed. + // (We can't exploit vertical or horizontal symmetry due to uncertain + // subpixel offsets. We could fix that by rotating xy offsets with the + // subpixel structure, but...no.) + const float3 w0 = eval_unorm_rgb_weights(xy_offset0, final_axis_importance); + const float3 w1 = eval_unorm_rgb_weights(xy_offset1, final_axis_importance); + const float3 w2 = eval_unorm_rgb_weights(xy_offset2, final_axis_importance); + const float3 w3 = eval_unorm_rgb_weights(xy_offset3, final_axis_importance); + const float3 w4 = eval_unorm_rgb_weights(xy_offset4, final_axis_importance); + const float3 w5 = eval_unorm_rgb_weights(xy_offset5, final_axis_importance); + const float3 w6 = eval_unorm_rgb_weights(xy_offset6, final_axis_importance); + const float3 w7 = eval_unorm_rgb_weights(xy_offset7, final_axis_importance); + const float3 w8 = w7.bgr; + const float3 w9 = w6.bgr; + const float3 w10 = w5.bgr; + const float3 w11 = w4.bgr; + const float3 w12 = w3.bgr; + const float3 w13 = w2.bgr; + const float3 w14 = w1.bgr; + const float3 w15 = w0.bgr; + // Get the weight sum to normalize the total to 1.0 later: + const float3 half_sum = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7; + const float3 w_sum = half_sum + half_sum.bgr; + const float3 w_sum_inv = float3(1.0)/(w_sum); + // Scale the pixel-space to texture offset matrix by the pixel diameter. + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + // Get uv sample offsets, taking advantage of row alignment: + const float2 uv_step_x = mul(true_pixel_to_tex_uv, float2(xy_step.x, 0.0)); + const float2 uv_step_y = mul(true_pixel_to_tex_uv, float2(0.0, xy_step.y)); + const float2 uv_offset0 = -1.5 * (uv_step_x + uv_step_y); + const float2 sample0_uv = tex_uv + uv_offset0; + const float2 sample4_uv = sample0_uv + uv_step_y; + const float2 sample8_uv = sample0_uv + uv_step_y * 2.0; + const float2 sample12_uv = sample0_uv + uv_step_y * 3.0; + // Load samples, linearizing if necessary, etc.: + const float3 sample0 = tex2Daa_tiled_linearize(tex, sample0_uv).rgb; + const float3 sample1 = tex2Daa_tiled_linearize(tex, sample0_uv + uv_step_x).rgb; + const float3 sample2 = tex2Daa_tiled_linearize(tex, sample0_uv + uv_step_x * 2.0).rgb; + const float3 sample3 = tex2Daa_tiled_linearize(tex, sample0_uv + uv_step_x * 3.0).rgb; + const float3 sample4 = tex2Daa_tiled_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2Daa_tiled_linearize(tex, sample4_uv + uv_step_x).rgb; + const float3 sample6 = tex2Daa_tiled_linearize(tex, sample4_uv + uv_step_x * 2.0).rgb; + const float3 sample7 = tex2Daa_tiled_linearize(tex, sample4_uv + uv_step_x * 3.0).rgb; + const float3 sample8 = tex2Daa_tiled_linearize(tex, sample8_uv).rgb; + const float3 sample9 = tex2Daa_tiled_linearize(tex, sample8_uv + uv_step_x).rgb; + const float3 sample10 = tex2Daa_tiled_linearize(tex, sample8_uv + uv_step_x * 2.0).rgb; + const float3 sample11 = tex2Daa_tiled_linearize(tex, sample8_uv + uv_step_x * 3.0).rgb; + const float3 sample12 = tex2Daa_tiled_linearize(tex, sample12_uv).rgb; + const float3 sample13 = tex2Daa_tiled_linearize(tex, sample12_uv + uv_step_x).rgb; + const float3 sample14 = tex2Daa_tiled_linearize(tex, sample12_uv + uv_step_x * 2.0).rgb; + const float3 sample15 = tex2Daa_tiled_linearize(tex, sample12_uv + uv_step_x * 3.0).rgb; + // Sum weighted samples (weight sum must equal 1.0 for each channel): + return w_sum_inv * ( + w0 * sample0 + w1 * sample1 + w2 * sample2 + w3 * sample3 + + w4 * sample4 + w5 * sample5 + w6 * sample6 + w7 * sample7 + + w8 * sample8 + w9 * sample9 + w10 * sample10 + w11 * sample11 + + w12 * sample12 + w13 * sample13 + w14 * sample14 + w15 * sample15); +} + +float3 tex2Daa_debug_dynamic(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ + // This function is for testing only: Use an NxN grid with dynamic weights. + static const int grid_size = 8; + assign_aa_cubic_constants(); + const float4 ssd_fai = get_subpixel_support_diam_and_final_axis_importance(); + const float2 subpixel_support_diameter = ssd_fai.xy; + const float2 final_axis_importance = ssd_fai.zw; + const float grid_radius_in_samples = (float(grid_size) - 1.0)/2.0; + const float2 filter_space_offset_step = + subpixel_support_diameter/float2(grid_size); + const float2 sample0_filter_space_offset = + -grid_radius_in_samples * filter_space_offset_step; + // Compute xy sample offsets and subpixel weights: + float3 weights[64]; //originally grid_size * grid_size + float3 weight_sum = float3(0.0, 0.0, 0.0); + for(int i = 0; i < grid_size; ++i) + { + for(int j = 0; j < grid_size; ++j) + { + // Weights based on xy distances: + const float2 offset = sample0_filter_space_offset + + float2(j, i) * filter_space_offset_step; + const float3 weight = eval_unorm_rgb_weights(offset, final_axis_importance); + weights[i*grid_size + j] = weight; + weight_sum += weight; + } + } + // Get uv offset vectors along x and y directions: + const float2x2 true_pixel_to_tex_uv = + float2x2(pixel_to_tex_uv * aa_pixel_diameter); + const float2 uv_offset_step_x = mul(true_pixel_to_tex_uv, + float2(filter_space_offset_step.x, 0.0)); + const float2 uv_offset_step_y = mul(true_pixel_to_tex_uv, + float2(0.0, filter_space_offset_step.y)); + // Get a starting sample location: + const float2 sample0_uv_offset = -grid_radius_in_samples * + (uv_offset_step_x + uv_offset_step_y); + const float2 sample0_uv = tex_uv + sample0_uv_offset; + // Load, weight, and sum [linearized] samples: + float3 sum = float3(0.0, 0.0, 0.0); + const float3 weight_sum_inv = float3(1.0)/weight_sum; + for(int i = 0; i < grid_size; ++i) + { + const float2 row_i_first_sample_uv = + sample0_uv + i * uv_offset_step_y; + for(int j = 0; j < grid_size; ++j) + { + const float2 sample_uv = + row_i_first_sample_uv + j * uv_offset_step_x; + sum += weights[i*grid_size + j] * + tex2Daa_tiled_linearize(tex, sample_uv).rgb; + } + } + return sum * weight_sum_inv; +} + + +/////////////////////// ANTIALIASING CODEPATH SELECTION ////////////////////// + +inline float3 tex2Daa(const sampler2D tex, const float2 tex_uv, + const float2x2 pixel_to_tex_uv, const float frame) +{ +#define DEBUG +#ifdef DEBUG + return tex2Daa_subpixel_weights_only( + tex, tex_uv, pixel_to_tex_uv); +#else + // Statically switch between antialiasing modes/levels: + return (aa_level < 0.5) ? tex2D_linearize(tex, tex_uv).rgb : + (aa_level < 3.5) ? tex2Daa_subpixel_weights_only( + tex, tex_uv, pixel_to_tex_uv) : + (aa_level < 4.5) ? tex2Daa4x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 5.5) ? tex2Daa5x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 6.5) ? tex2Daa6x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 7.5) ? tex2Daa7x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 11.5) ? tex2Daa8x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 15.5) ? tex2Daa12x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 19.5) ? tex2Daa16x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 23.5) ? tex2Daa20x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 253.5) ? tex2Daa24x(tex, tex_uv, pixel_to_tex_uv, frame) : + (aa_level < 254.5) ? tex2Daa_debug_16x_regular( + tex, tex_uv, pixel_to_tex_uv, frame) : + tex2Daa_debug_dynamic(tex, tex_uv, pixel_to_tex_uv, frame); +#endif +} + + +#endif // TEX2DANTIALIAS_H + +///////////////////////// END TEX2DANTIALIAS ///////////////////////// + +//#include "geometry-functions.h" + +///////////////////////// BEGIN GEOMETRY-FUNCTIONS ///////////////////////// + +#ifndef GEOMETRY_FUNCTIONS_H +#define GEOMETRY_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// already included elsewhere +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" +//#include "bind-shader-h" + + +//////////////////////////// MACROS AND CONSTANTS //////////////////////////// + +// Curvature-related constants: +#define MAX_POINT_CLOUD_SIZE 9 + + +///////////////////////////// CURVATURE FUNCTIONS ///////////////////////////// + +float2 quadratic_solve(const float a, const float b_over_2, const float c) +{ + // Requires: 1.) a, b, and c are quadratic formula coefficients + // 2.) b_over_2 = b/2.0 (simplifies terms to factor 2 out) + // 3.) b_over_2 must be guaranteed < 0.0 (avoids a branch) + // Returns: Returns float2(first_solution, discriminant), so the caller + // can choose how to handle the "no intersection" case. The + // Kahan or Citardauq formula is used for numerical robustness. + const float discriminant = b_over_2*b_over_2 - a*c; + const float solution0 = c/(-b_over_2 + sqrt(discriminant)); + return float2(solution0, discriminant); +} + +float2 intersect_sphere(const float3 view_vec, const float3 eye_pos_vec) +{ + // Requires: 1.) view_vec and eye_pos_vec are 3D vectors in the sphere's + // local coordinate frame (eye_pos_vec is a position, i.e. + // a vector from the origin to the eye/camera) + // 2.) geom_radius is a global containing the sphere's radius + // Returns: Cast a ray of direction view_vec from eye_pos_vec at a + // sphere of radius geom_radius, and return the distance to + // the first intersection in units of length(view_vec). + // http://wiki.cgsociety.org/index.php/Ray_Sphere_Intersection + // Quadratic formula coefficients (b_over_2 is guaranteed negative): + const float a = dot(view_vec, view_vec); + const float b_over_2 = dot(view_vec, eye_pos_vec); // * 2.0 factored out + const float c = dot(eye_pos_vec, eye_pos_vec) - geom_radius*geom_radius; + return quadratic_solve(a, b_over_2, c); +} + +float2 intersect_cylinder(const float3 view_vec, const float3 eye_pos_vec) +{ + // Requires: 1.) view_vec and eye_pos_vec are 3D vectors in the sphere's + // local coordinate frame (eye_pos_vec is a position, i.e. + // a vector from the origin to the eye/camera) + // 2.) geom_radius is a global containing the cylinder's radius + // Returns: Cast a ray of direction view_vec from eye_pos_vec at a + // cylinder of radius geom_radius, and return the distance to + // the first intersection in units of length(view_vec). The + // derivation of the coefficients is in Christer Ericson's + // Real-Time Collision Detection, p. 195-196, and this version + // uses LaGrange's identity to reduce operations. + // Arbitrary "cylinder top" reference point for an infinite cylinder: + const float3 cylinder_top_vec = float3(0.0, geom_radius, 0.0); + const float3 cylinder_axis_vec = float3(0.0, 1.0, 0.0);//float3(0.0, 2.0*geom_radius, 0.0); + const float3 top_to_eye_vec = eye_pos_vec - cylinder_top_vec; + const float3 axis_x_view = cross(cylinder_axis_vec, view_vec); + const float3 axis_x_top_to_eye = cross(cylinder_axis_vec, top_to_eye_vec); + // Quadratic formula coefficients (b_over_2 is guaranteed negative): + const float a = dot(axis_x_view, axis_x_view); + const float b_over_2 = dot(axis_x_top_to_eye, axis_x_view); + const float c = dot(axis_x_top_to_eye, axis_x_top_to_eye) - + geom_radius*geom_radius;//*dot(cylinder_axis_vec, cylinder_axis_vec); + return quadratic_solve(a, b_over_2, c); +} + +float2 cylinder_xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect) +{ + // Requires: An xyz intersection position on a cylinder. + // Returns: video_uv coords mapped to range [-0.5, 0.5] + // Mapping: Define square_uv.x to be the signed arc length in xz-space, + // and define square_uv.y = -intersection_pos_local.y (+v = -y). + // Start with a numerically robust arc length calculation. + const float angle_from_image_center = atan2(intersection_pos_local.x, + intersection_pos_local.z); + const float signed_arc_len = angle_from_image_center * geom_radius; + // Get a uv-mapping where [-0.5, 0.5] maps to a "square" area, then divide + // by the aspect ratio to stretch the mapping appropriately: + const float2 square_uv = float2(signed_arc_len, -intersection_pos_local.y); + const float2 video_uv = square_uv / geom_aspect; + return video_uv; +} + +float3 cylinder_uv_to_xyz(const float2 video_uv, const float2 geom_aspect) +{ + // Requires: video_uv coords mapped to range [-0.5, 0.5] + // Returns: An xyz intersection position on a cylinder. This is the + // inverse of cylinder_xyz_to_uv(). + // Expand video_uv by the aspect ratio to get proportionate x/y lengths, + // then calculate an xyz position for the cylindrical mapping above. + const float2 square_uv = video_uv * geom_aspect; + const float arc_len = square_uv.x; + const float angle_from_image_center = arc_len / geom_radius; + const float x_pos = sin(angle_from_image_center) * geom_radius; + const float z_pos = cos(angle_from_image_center) * geom_radius; + // Or: z = sqrt(geom_radius**2 - x**2) + // Or: z = geom_radius/sqrt(1.0 + tan(angle)**2), x = z * tan(angle) + const float3 intersection_pos_local = float3(x_pos, -square_uv.y, z_pos); + return intersection_pos_local; +} + +float2 sphere_xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect) +{ + // Requires: An xyz intersection position on a sphere. + // Returns: video_uv coords mapped to range [-0.5, 0.5] + // Mapping: First define square_uv.x/square_uv.y == + // intersection_pos_local.x/intersection_pos_local.y. Then, + // length(square_uv) is the arc length from the image center + // at (0.0, 0.0, geom_radius) along the tangent great circle. + // Credit for this mapping goes to cgwg: I never managed to + // understand his code, but he told me his mapping was based on + // great circle distances when I asked him about it, which + // informed this very similar (almost identical) mapping. + // Start with a numerically robust arc length calculation between the ray- + // sphere intersection point and the image center using a method posted by + // Roger Stafford on comp.soft-sys.matlab: + // https://groups.google.com/d/msg/comp.soft-sys.matlab/zNbUui3bjcA/c0HV_bHSx9cJ + const float3 image_center_pos_local = float3(0.0, 0.0, geom_radius); + const float cp_len = + length(cross(intersection_pos_local, image_center_pos_local)); + const float dp = dot(intersection_pos_local, image_center_pos_local); + const float angle_from_image_center = atan2(cp_len, dp); + const float arc_len = angle_from_image_center * geom_radius; + // Get a uv-mapping where [-0.5, 0.5] maps to a "square" area, then divide + // by the aspect ratio to stretch the mapping appropriately: + const float2 square_uv_unit = normalize(float2(intersection_pos_local.x, + -intersection_pos_local.y)); + const float2 square_uv = arc_len * square_uv_unit; + const float2 video_uv = square_uv / geom_aspect; + return video_uv; +} + +float3 sphere_uv_to_xyz(const float2 video_uv, const float2 geom_aspect) +{ + // Requires: video_uv coords mapped to range [-0.5, 0.5] + // Returns: An xyz intersection position on a sphere. This is the + // inverse of sphere_xyz_to_uv(). + // Expand video_uv by the aspect ratio to get proportionate x/y lengths, + // then calculate an xyz position for the spherical mapping above. + const float2 square_uv = video_uv * geom_aspect; + // Using length or sqrt here butchers the framerate on my 8800GTS if + // this function is called too many times, and so does taking the max + // component of square_uv/square_uv_unit (program length threshold?). + //float arc_len = length(square_uv); + const float2 square_uv_unit = normalize(square_uv); + const float arc_len = square_uv.y/square_uv_unit.y; + const float angle_from_image_center = arc_len / geom_radius; + const float xy_dist_from_sphere_center = + sin(angle_from_image_center) * geom_radius; + //float2 xy_pos = xy_dist_from_sphere_center * (square_uv/FIX_ZERO(arc_len)); + const float2 xy_pos = xy_dist_from_sphere_center * square_uv_unit; + const float z_pos = cos(angle_from_image_center) * geom_radius; + const float3 intersection_pos_local = float3(xy_pos.x, -xy_pos.y, z_pos); + return intersection_pos_local; +} + +float2 sphere_alt_xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect) +{ + // Requires: An xyz intersection position on a cylinder. + // Returns: video_uv coords mapped to range [-0.5, 0.5] + // Mapping: Define square_uv.x to be the signed arc length in xz-space, + // and define square_uv.y == signed arc length in yz-space. + // See cylinder_xyz_to_uv() for implementation details (very similar). + const float2 angle_from_image_center = atan2( + float2(intersection_pos_local.x, -intersection_pos_local.y), + intersection_pos_local.zz); + const float2 signed_arc_len = angle_from_image_center * geom_radius; + const float2 video_uv = signed_arc_len / geom_aspect; + return video_uv; +} + +float3 sphere_alt_uv_to_xyz(const float2 video_uv, const float2 geom_aspect) +{ + // Requires: video_uv coords mapped to range [-0.5, 0.5] + // Returns: An xyz intersection position on a sphere. This is the + // inverse of sphere_alt_xyz_to_uv(). + // See cylinder_uv_to_xyz() for implementation details (very similar). + const float2 square_uv = video_uv * geom_aspect; + const float2 arc_len = square_uv; + const float2 angle_from_image_center = arc_len / geom_radius; + const float2 xy_pos = sin(angle_from_image_center) * geom_radius; + const float z_pos = sqrt(geom_radius*geom_radius - dot(xy_pos, xy_pos)); + return float3(xy_pos.x, -xy_pos.y, z_pos); +} + +inline float2 intersect(const float3 view_vec_local, const float3 eye_pos_local, + const float geom_mode) +{ + return geom_mode < 2.5 ? intersect_sphere(view_vec_local, eye_pos_local) : + intersect_cylinder(view_vec_local, eye_pos_local); +} + +inline float2 xyz_to_uv(const float3 intersection_pos_local, + const float2 geom_aspect, const float geom_mode) +{ + return geom_mode < 1.5 ? + sphere_xyz_to_uv(intersection_pos_local, geom_aspect) : + geom_mode < 2.5 ? + sphere_alt_xyz_to_uv(intersection_pos_local, geom_aspect) : + cylinder_xyz_to_uv(intersection_pos_local, geom_aspect); +} + +inline float3 uv_to_xyz(const float2 uv, const float2 geom_aspect, + const float geom_mode) +{ + return geom_mode < 1.5 ? sphere_uv_to_xyz(uv, geom_aspect) : + geom_mode < 2.5 ? sphere_alt_uv_to_xyz(uv, geom_aspect) : + cylinder_uv_to_xyz(uv, geom_aspect); +} + +float2 view_vec_to_uv(const float3 view_vec_local, const float3 eye_pos_local, + const float2 geom_aspect, const float geom_mode, out float3 intersection_pos) +{ + // Get the intersection point on the primitive, given an eye position + // and view vector already in its local coordinate frame: + const float2 intersect_dist_and_discriminant = intersect(view_vec_local, + eye_pos_local, geom_mode); + const float3 intersection_pos_local = eye_pos_local + + view_vec_local * intersect_dist_and_discriminant.x; + // Save the intersection position to an output parameter: + intersection_pos = intersection_pos_local; + // Transform into uv coords, but give out-of-range coords if the + // view ray doesn't intersect the primitive in the first place: + return intersect_dist_and_discriminant.y > 0.005 ? + xyz_to_uv(intersection_pos_local, geom_aspect, geom_mode) : float2(1.0); +} + +float3 get_ideal_global_eye_pos_for_points(float3 eye_pos, + const float2 geom_aspect, const float3 global_coords[MAX_POINT_CLOUD_SIZE], + const int num_points) +{ + // Requires: Parameters: + // 1.) Starting eye_pos is a global 3D position at which the + // camera contains all points in global_coords[] in its FOV + // 2.) geom_aspect = get_aspect_vector( + // output_size.x / output_size.y); + // 3.) global_coords is a point cloud containing global xyz + // coords of extreme points on the simulated CRT screen. + // Globals: + // 1.) geom_view_dist must be > 0.0. It controls the "near + // plane" used to interpret flat_video_uv as a view + // vector, which controls the field of view (FOV). + // Eyespace coordinate frame: +x = right, +y = up, +z = back + // Returns: Return an eye position at which the point cloud spans as + // much of the screen as possible (given the FOV controlled by + // geom_view_dist) without being cropped or sheared. + // Algorithm: + // 1.) Move the eye laterally to a point which attempts to maximize the + // the amount we can move forward without clipping the CRT screen. + // 2.) Move forward by as much as possible without clipping the CRT. + // Get the allowed movement range by solving for the eye_pos offsets + // that result in each point being projected to a screen edge/corner in + // pseudo-normalized device coords (where xy ranges from [-0.5, 0.5] + // and z = eyespace z): + // pndc_coord = float3(float2(eyespace_xyz.x, -eyespace_xyz.y)* + // geom_view_dist / (geom_aspect * -eyespace_xyz.z), eyespace_xyz.z); + // Notes: + // The field of view is controlled by geom_view_dist's magnitude relative to + // the view vector's x and y components: + // view_vec.xy ranges from [-0.5, 0.5] * geom_aspect + // view_vec.z = -geom_view_dist + // But for the purposes of perspective divide, it should be considered: + // view_vec.xy ranges from [-0.5, 0.5] * geom_aspect / geom_view_dist + // view_vec.z = -1.0 + static const int max_centering_iters = 1; // Keep for easy testing. + for(int iter = 0; iter < max_centering_iters; iter++) + { + // 0.) Get the eyespace coordinates of our point cloud: + float3 eyespace_coords[MAX_POINT_CLOUD_SIZE]; + for(int i = 0; i < num_points; i++) + { + eyespace_coords[i] = global_coords[i] - eye_pos; + } + // 1a.)For each point, find out how far we can move eye_pos in each + // lateral direction without the point clipping the frustum. + // Eyespace +y = up, screenspace +y = down, so flip y after + // applying the eyespace offset (on the way to "clip space"). + // Solve for two offsets per point based on: + // (eyespace_xyz.xy - offset_dr) * float2(1.0, -1.0) * + // geom_view_dist / (geom_aspect * -eyespace_xyz.z) = float2(-0.5) + // (eyespace_xyz.xy - offset_dr) * float2(1.0, -1.0) * + // geom_view_dist / (geom_aspect * -eyespace_xyz.z) = float2(0.5) + // offset_ul and offset_dr represent the farthest we can move the + // eye_pos up-left and down-right. Save the min of all offset_dr's + // and the max of all offset_ul's (since it's negative). + float abs_radius = abs(geom_radius); // In case anyone gets ideas. ;) + float2 offset_dr_min = float2(10.0 * abs_radius, 10.0 * abs_radius); + float2 offset_ul_max = float2(-10.0 * abs_radius, -10.0 * abs_radius); + for(int i = 0; i < num_points; i++) + { + static const float2 flipy = float2(1.0, -1.0); + float3 eyespace_xyz = eyespace_coords[i]; + float2 offset_dr = eyespace_xyz.xy - float2(-0.5) * + (geom_aspect * -eyespace_xyz.z) / (geom_view_dist * flipy); + float2 offset_ul = eyespace_xyz.xy - float2(0.5) * + (geom_aspect * -eyespace_xyz.z) / (geom_view_dist * flipy); + offset_dr_min = min(offset_dr_min, offset_dr); + offset_ul_max = max(offset_ul_max, offset_ul); + } + // 1b.)Update eye_pos: Adding the average of offset_ul_max and + // offset_dr_min gives it equal leeway on the top vs. bottom + // and left vs. right. Recalculate eyespace_coords accordingly. + float2 center_offset = 0.5 * (offset_ul_max + offset_dr_min); + eye_pos.xy += center_offset; + for(int i = 0; i < num_points; i++) + { + eyespace_coords[i] = global_coords[i] - eye_pos; + } + // 2a.)For each point, find out how far we can move eye_pos forward + // without the point clipping the frustum. Flip the y + // direction in advance (matters for a later step, not here). + // Solve for four offsets per point based on: + // eyespace_xyz_flipy.x * geom_view_dist / + // (geom_aspect.x * (offset_z - eyespace_xyz_flipy.z)) =-0.5 + // eyespace_xyz_flipy.y * geom_view_dist / + // (geom_aspect.y * (offset_z - eyespace_xyz_flipy.z)) =-0.5 + // eyespace_xyz_flipy.x * geom_view_dist / + // (geom_aspect.x * (offset_z - eyespace_xyz_flipy.z)) = 0.5 + // eyespace_xyz_flipy.y * geom_view_dist / + // (geom_aspect.y * (offset_z - eyespace_xyz_flipy.z)) = 0.5 + // We'll vectorize the actual computation. Take the maximum of + // these four for a single offset, and continue taking the max + // for every point (use max because offset.z is negative). + float offset_z_max = -10.0 * geom_radius * geom_view_dist; + for(int i = 0; i < num_points; i++) + { + float3 eyespace_xyz_flipy = eyespace_coords[i] * + float3(1.0, -1.0, 1.0); + float4 offset_zzzz = eyespace_xyz_flipy.zzzz + + (eyespace_xyz_flipy.xyxy * geom_view_dist) / + (float4(-0.5, -0.5, 0.5, 0.5) * float4(geom_aspect, geom_aspect)); + // Ignore offsets that push positive x/y values to opposite + // boundaries, and vice versa, and don't let the camera move + // past a point in the dead center of the screen: + offset_z_max = (eyespace_xyz_flipy.x < 0.0) ? + max(offset_z_max, offset_zzzz.x) : offset_z_max; + offset_z_max = (eyespace_xyz_flipy.y < 0.0) ? + max(offset_z_max, offset_zzzz.y) : offset_z_max; + offset_z_max = (eyespace_xyz_flipy.x > 0.0) ? + max(offset_z_max, offset_zzzz.z) : offset_z_max; + offset_z_max = (eyespace_xyz_flipy.y > 0.0) ? + max(offset_z_max, offset_zzzz.w) : offset_z_max; + offset_z_max = max(offset_z_max, eyespace_xyz_flipy.z); + } + // 2b.)Update eye_pos: Add the maximum (smallest negative) z offset. + eye_pos.z += offset_z_max; + } + return eye_pos; +} + +float3 get_ideal_global_eye_pos(const float3x3 local_to_global, + const float2 geom_aspect, const float geom_mode) +{ + // Start with an initial eye_pos that includes the entire primitive + // (sphere or cylinder) in its field-of-view: + const float3 high_view = float3(0.0, geom_aspect.y, -geom_view_dist); + const float3 low_view = high_view * float3(1.0, -1.0, 1.0); + const float len_sq = dot(high_view, high_view); + const float fov = abs(acos(dot(high_view, low_view)/len_sq)); + // Trigonometry/similar triangles say distance = geom_radius/sin(fov/2): + const float eye_z_spherical = geom_radius/sin(fov*0.5); + const float3 eye_pos = geom_mode < 2.5 ? + float3(0.0, 0.0, eye_z_spherical) : + float3(0.0, 0.0, max(geom_view_dist, eye_z_spherical)); + + // Get global xyz coords of extreme sample points on the simulated CRT + // screen. Start with the center, edge centers, and corners of the + // video image. We can't ignore backfacing points: They're occluded + // by closer points on the primitive, but they may NOT be occluded by + // the convex hull of the remaining samples (i.e. the remaining convex + // hull might not envelope points that do occlude a back-facing point.) + static const int num_points = MAX_POINT_CLOUD_SIZE; + float3 global_coords[MAX_POINT_CLOUD_SIZE]; + global_coords[0] = mul(local_to_global, uv_to_xyz(float2(0.0, 0.0), geom_aspect, geom_mode)); + global_coords[1] = mul(local_to_global, uv_to_xyz(float2(0.0, -0.5), geom_aspect, geom_mode)); + global_coords[2] = mul(local_to_global, uv_to_xyz(float2(0.0, 0.5), geom_aspect, geom_mode)); + global_coords[3] = mul(local_to_global, uv_to_xyz(float2(-0.5, 0.0), geom_aspect, geom_mode)); + global_coords[4] = mul(local_to_global, uv_to_xyz(float2(0.5, 0.0), geom_aspect, geom_mode)); + global_coords[5] = mul(local_to_global, uv_to_xyz(float2(-0.5, -0.5), geom_aspect, geom_mode)); + global_coords[6] = mul(local_to_global, uv_to_xyz(float2(0.5, -0.5), geom_aspect, geom_mode)); + global_coords[7] = mul(local_to_global, uv_to_xyz(float2(-0.5, 0.5), geom_aspect, geom_mode)); + global_coords[8] = mul(local_to_global, uv_to_xyz(float2(0.5, 0.5), geom_aspect, geom_mode)); + // Adding more inner image points could help in extreme cases, but too many + // points will kille the framerate. For safety, default to the initial + // eye_pos if any z coords are negative: + float num_negative_z_coords = 0.0; + for(int i = 0; i < num_points; i++) + { + num_negative_z_coords += float(global_coords[0].z < 0.0); + } + // Outsource the optimized eye_pos calculation: + return num_negative_z_coords > 0.5 ? eye_pos : + get_ideal_global_eye_pos_for_points(eye_pos, geom_aspect, + global_coords, num_points); +} + +float3x3 get_pixel_to_object_matrix(const float3x3 global_to_local, + const float3 eye_pos_local, const float3 view_vec_global, + const float3 intersection_pos_local, const float3 normal, + const float2 output_size_inv) +{ + // Requires: See get_curved_video_uv_coords_and_tangent_matrix for + // descriptions of each parameter. + // Returns: Return a transformation matrix from 2D pixel-space vectors + // (where (+1.0, +1.0) is a vector to one pixel down-right, + // i.e. same directionality as uv texels) to 3D object-space + // vectors in the CRT's local coordinate frame (right-handed) + // ***which are tangent to the CRT surface at the intersection + // position.*** (Basically, we want to convert pixel-space + // vectors to 3D vectors along the CRT's surface, for later + // conversion to uv vectors.) + // Shorthand inputs: + const float3 pos = intersection_pos_local; + const float3 eye_pos = eye_pos_local; + // Get a piecewise-linear matrix transforming from "pixelspace" offset + // vectors (1.0 = one pixel) to object space vectors in the tangent + // plane (faster than finding 3 view-object intersections). + // 1.) Get the local view vecs for the pixels to the right and down: + const float3 view_vec_right_global = view_vec_global + + float3(output_size_inv.x, 0.0, 0.0); + const float3 view_vec_down_global = view_vec_global + + float3(0.0, -output_size_inv.y, 0.0); + const float3 view_vec_right_local = + mul(global_to_local, view_vec_right_global); + const float3 view_vec_down_local = + mul(global_to_local, view_vec_down_global); + // 2.) Using the true intersection point, intersect the neighboring + // view vectors with the tangent plane: + const float3 intersection_vec_dot_normal = float3(dot(pos - eye_pos, normal), dot(pos - eye_pos, normal), dot(pos - eye_pos, normal)); + const float3 right_pos = eye_pos + (intersection_vec_dot_normal / + dot(view_vec_right_local, normal))*view_vec_right_local; + const float3 down_pos = eye_pos + (intersection_vec_dot_normal / + dot(view_vec_down_local, normal))*view_vec_down_local; + // 3.) Subtract the original intersection pos from its neighbors; the + // resulting vectors are object-space vectors tangent to the plane. + // These vectors are the object-space transformations of (1.0, 0.0) + // and (0.0, 1.0) pixel offsets, so they form the first two basis + // vectors of a pixelspace to object space transformation. This + // transformation is 2D to 3D, so use (0, 0, 0) for the third vector. + const float3 object_right_vec = right_pos - pos; + const float3 object_down_vec = down_pos - pos; + const float3x3 pixel_to_object = float3x3( + object_right_vec.x, object_down_vec.x, 0.0, + object_right_vec.y, object_down_vec.y, 0.0, + object_right_vec.z, object_down_vec.z, 0.0); + return pixel_to_object; +} + +float3x3 get_object_to_tangent_matrix(const float3 intersection_pos_local, + const float3 normal, const float2 geom_aspect, const float geom_mode) +{ + // Requires: See get_curved_video_uv_coords_and_tangent_matrix for + // descriptions of each parameter. + // Returns: Return a transformation matrix from 3D object-space vectors + // in the CRT's local coordinate frame (right-handed, +y = up) + // to 2D video_uv vectors (+v = down). + // Description: + // The TBN matrix formed by the [tangent, bitangent, normal] basis + // vectors transforms ordinary vectors from tangent->object space. + // The cotangent matrix formed by the [cotangent, cobitangent, normal] + // basis vectors transforms normal vectors (covectors) from + // tangent->object space. It's the inverse-transpose of the TBN matrix. + // We want the inverse of the TBN matrix (transpose of the cotangent + // matrix), which transforms ordinary vectors from object->tangent space. + // Start by calculating the relevant basis vectors in accordance with + // Christian Schüler's blog post "Followup: Normal Mapping Without + // Precomputed Tangents": http://www.thetenthplanet.de/archives/1180 + // With our particular uv mapping, the scale of the u and v directions + // is determined entirely by the aspect ratio for cylindrical and ordinary + // spherical mappings, and so tangent and bitangent lengths are also + // determined by it (the alternate mapping is more complex). Therefore, we + // must ensure appropriate cotangent and cobitangent lengths as well. + // Base these off the uv<=>xyz mappings for each primitive. + const float3 pos = intersection_pos_local; + static const float3 x_vec = float3(1.0, 0.0, 0.0); + static const float3 y_vec = float3(0.0, 1.0, 0.0); + // The tangent and bitangent vectors correspond with increasing u and v, + // respectively. Mathematically we'd base the cotangent/cobitangent on + // those, but we'll compute the cotangent/cobitangent directly when we can. + float3 cotangent_unscaled, cobitangent_unscaled; + // geom_mode should be constant-folded without RUNTIME_GEOMETRY_MODE. + if(geom_mode < 1.5) + { + // Sphere: + // tangent = normalize(cross(normal, cross(x_vec, pos))) * geom_aspect.x + // bitangent = normalize(cross(cross(y_vec, pos), normal)) * geom_aspect.y + // inv_determinant = 1.0/length(cross(bitangent, tangent)) + // cotangent = cross(normal, bitangent) * inv_determinant + // == normalize(cross(y_vec, pos)) * geom_aspect.y * inv_determinant + // cobitangent = cross(tangent, normal) * inv_determinant + // == normalize(cross(x_vec, pos)) * geom_aspect.x * inv_determinant + // Simplified (scale by inv_determinant below): + cotangent_unscaled = normalize(cross(y_vec, pos)) * geom_aspect.y; + cobitangent_unscaled = normalize(cross(x_vec, pos)) * geom_aspect.x; + } + else if(geom_mode < 2.5) + { + // Sphere, alternate mapping: + // This mapping works a bit like the cylindrical mapping in two + // directions, which makes the lengths and directions more complex. + // Unfortunately, I can't find much of a shortcut: + const float3 tangent = normalize( + cross(y_vec, float3(pos.x, 0.0, pos.z))) * geom_aspect.x; + const float3 bitangent = normalize( + cross(x_vec, float3(0.0, pos.yz))) * geom_aspect.y; + cotangent_unscaled = cross(normal, bitangent); + cobitangent_unscaled = cross(tangent, normal); + } + else + { + // Cylinder: + // tangent = normalize(cross(y_vec, normal)) * geom_aspect.x; + // bitangent = float3(0.0, -geom_aspect.y, 0.0); + // inv_determinant = 1.0/length(cross(bitangent, tangent)) + // cotangent = cross(normal, bitangent) * inv_determinant + // == normalize(cross(y_vec, pos)) * geom_aspect.y * inv_determinant + // cobitangent = cross(tangent, normal) * inv_determinant + // == float3(0.0, -geom_aspect.x, 0.0) * inv_determinant + cotangent_unscaled = cross(y_vec, normal) * geom_aspect.y; + cobitangent_unscaled = float3(0.0, -geom_aspect.x, 0.0); + } + const float3 computed_normal = + cross(cobitangent_unscaled, cotangent_unscaled); + const float inv_determinant = rsqrt(dot(computed_normal, computed_normal)); + const float3 cotangent = cotangent_unscaled * inv_determinant; + const float3 cobitangent = cobitangent_unscaled * inv_determinant; + // The [cotangent, cobitangent, normal] column vecs form the cotangent + // frame, i.e. the inverse-transpose TBN matrix. Get its transpose: + const float3x3 object_to_tangent = float3x3(cotangent, cobitangent, normal); + return object_to_tangent; +} + +float2 get_curved_video_uv_coords_and_tangent_matrix( + const float2 flat_video_uv, const float3 eye_pos_local, + const float2 output_size_inv, const float2 geom_aspect, + const float geom_mode, const float3x3 global_to_local, + out float2x2 pixel_to_tangent_video_uv) +{ + // Requires: Parameters: + // 1.) flat_video_uv coords are in range [0.0, 1.0], where + // (0.0, 0.0) is the top-left corner of the screen and + // (1.0, 1.0) is the bottom-right corner. + // 2.) eye_pos_local is the 3D camera position in the simulated + // CRT's local coordinate frame. For best results, it must + // be computed based on the same geom_view_dist used here. + // 3.) output_size_inv = float2(1.0)/output_size + // 4.) geom_aspect = get_aspect_vector( + // output_size.x / output_size.y); + // 5.) geom_mode is a static or runtime mode setting: + // 0 = off, 1 = sphere, 2 = sphere alt., 3 = cylinder + // 6.) global_to_local is a 3x3 matrix transforming (ordinary) + // worldspace vectors to the CRT's local coordinate frame + // Globals: + // 1.) geom_view_dist must be > 0.0. It controls the "near + // plane" used to interpret flat_video_uv as a view + // vector, which controls the field of view (FOV). + // Returns: Return final uv coords in [0.0, 1.0], and return a pixel- + // space to video_uv tangent-space matrix in the out parameter. + // (This matrix assumes pixel-space +y = down, like +v = down.) + // We'll transform flat_video_uv into a view vector, project + // the view vector from the camera/eye, intersect with a sphere + // or cylinder representing the simulated CRT, and convert the + // intersection position into final uv coords and a local + // transformation matrix. + // First get the 3D view vector (geom_aspect and geom_view_dist are globals): + // 1.) Center uv around (0.0, 0.0) and make (-0.5, -0.5) and (0.5, 0.5) + // correspond to the top-left/bottom-right output screen corners. + // 2.) Multiply by geom_aspect to preemptively "undo" Retroarch's screen- + // space 2D aspect correction. We'll reapply it in uv-space. + // 3.) (x, y) = (u, -v), because +v is down in 2D screenspace, but +y + // is up in 3D worldspace (enforce a right-handed system). + // 4.) The view vector z controls the "near plane" distance and FOV. + // For the effect of "looking through a window" at a CRT, it should be + // set equal to the user's distance from their physical screen, in + // units of the viewport's physical diagonal size. + const float2 view_uv = (flat_video_uv - float2(0.5)) * geom_aspect; + const float3 view_vec_global = + float3(view_uv.x, -view_uv.y, -geom_view_dist); + // Transform the view vector into the CRT's local coordinate frame, convert + // to video_uv coords, and get the local 3D intersection position: + const float3 view_vec_local = mul(global_to_local, view_vec_global); + float3 pos; + const float2 centered_uv = view_vec_to_uv( + view_vec_local, eye_pos_local, geom_aspect, geom_mode, pos); + const float2 video_uv = centered_uv + float2(0.5); + // Get a pixel-to-tangent-video-uv matrix. The caller could deal with + // all but one of these cases, but that would be more complicated. + #ifdef DRIVERS_ALLOW_DERIVATIVES + // Derivatives obtain a matrix very fast, but the direction of pixel- + // space +y seems to depend on the pass. Enforce the correct direction + // on a best-effort basis (but it shouldn't matter for antialiasing). + const float2 duv_dx = ddx(video_uv); + const float2 duv_dy = ddy(video_uv); + #ifdef LAST_PASS + pixel_to_tangent_video_uv = float2x2( + duv_dx.x, duv_dy.x, + -duv_dx.y, -duv_dy.y); + #else + pixel_to_tangent_video_uv = float2x2( + duv_dx.x, duv_dy.x, + duv_dx.y, duv_dy.y); + #endif + #else + // Manually define a transformation matrix. We'll assume pixel-space + // +y = down, just like +v = down. + if(geom_force_correct_tangent_matrix) + { + // Get the surface normal based on the local intersection position: + const float3 normal_base = geom_mode < 2.5 ? pos : + float3(pos.x, 0.0, pos.z); + const float3 normal = normalize(normal_base); + // Get pixel-to-object and object-to-tangent matrices and combine + // them into a 2x2 pixel-to-tangent matrix for video_uv offsets: + const float3x3 pixel_to_object = get_pixel_to_object_matrix( + global_to_local, eye_pos_local, view_vec_global, pos, normal, + output_size_inv); + const float3x3 object_to_tangent = get_object_to_tangent_matrix( + pos, normal, geom_aspect, geom_mode); + const float3x3 pixel_to_tangent3x3 = + mul(object_to_tangent, pixel_to_object); + pixel_to_tangent_video_uv = float2x2( + pixel_to_tangent3x3[0][0], pixel_to_tangent3x3[0][1], pixel_to_tangent3x3[1][0], pixel_to_tangent3x3[1][1]);//._m00_m01_m10_m11); //TODO/FIXME: needs to correct for column-major?? + } + else + { + // Ignore curvature, and just consider flat scaling. The + // difference is only apparent with strong curvature: + pixel_to_tangent_video_uv = float2x2( + output_size_inv.x, 0.0, 0.0, output_size_inv.y); + } + #endif + return video_uv; +} + +float get_border_dim_factor(const float2 video_uv, const float2 geom_aspect) +{ + // COPYRIGHT NOTE FOR THIS FUNCTION: + // Copyright (C) 2010-2012 cgwg, 2014 TroggleMonkey + // This function uses an algorithm first coded in several of cgwg's GPL- + // licensed lines in crt-geom-curved.cg and its ancestors. The line + // between algorithm and code is nearly indistinguishable here, so it's + // unclear whether I could even release this project under a non-GPL + // license with this function included. + + // Calculate border_dim_factor from the proximity to uv-space image + // borders; geom_aspect/border_size/border/darkness/border_compress are globals: + const float2 edge_dists = min(video_uv, float2(1.0) - video_uv) * + geom_aspect; + const float2 border_penetration = + max(float2(border_size) - edge_dists, float2(0.0)); + const float penetration_ratio = length(border_penetration)/border_size; + const float border_escape_ratio = max(1.0 - penetration_ratio, 0.0); + const float border_dim_factor = + pow(border_escape_ratio, border_darkness) * max(1.0, border_compress); + return min(border_dim_factor, 1.0); +} + + + +#endif // GEOMETRY_FUNCTIONS_H + +///////////////////////// END GEOMETRY-FUNCTIONS ///////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +float2x2 mul_scale(float2 scale, float2x2 matrix) +{ + //float2x2 scale_matrix = float2x2(scale.x, 0.0, 0.0, scale.y); + //return mul(scale_matrix, matrix); + float4 intermed = float4(matrix[0][0],matrix[0][1],matrix[1][0],matrix[1][1]) * scale.xxyy; + return float2x2(intermed.x, intermed.y, intermed.z, intermed.w); +} + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + gl_Position = position; + vTexCoord = texCoord * 1.0001; + tex_uv = vTexCoord.xy; + video_and_texture_size_inv = + float4(1.0, 1.0, 1.0, 1.0) / float4(video_size, texture_size); + output_size_inv = float2(1.0, 1.0)/output_size; + + // Get aspect/overscan vectors from scalar parameters (likely uniforms): + const float viewport_aspect_ratio = output_size.x/output_size.y; + const float2 geom_aspect = get_aspect_vector(viewport_aspect_ratio); + const float2 geom_overscan = get_geom_overscan_vector(); + geom_aspect_and_overscan = float4(geom_aspect, geom_overscan); + + #ifdef RUNTIME_GEOMETRY_TILT + // Create a local-to-global rotation matrix for the CRT's coordinate + // frame and its global-to-local inverse. Rotate around the x axis + // first (pitch) and then the y axis (yaw) with yucky Euler angles. + // Positive angles go clockwise around the right-vec and up-vec. + // Runtime shader parameters prevent us from computing these globally, + // but we can still combine the pitch/yaw matrices by hand to cut a + // few instructions. Note that cg matrices fill row1 first, then row2, + // etc. (row-major order). + const float2 geom_tilt_angle = get_geom_tilt_angle_vector(); + const float2 sin_tilt = sin(geom_tilt_angle); + const float2 cos_tilt = cos(geom_tilt_angle); + // Conceptual breakdown: + static const float3x3 rot_x_matrix = float3x3( + 1.0, 0.0, 0.0, + 0.0, cos_tilt.y, -sin_tilt.y, + 0.0, sin_tilt.y, cos_tilt.y); + static const float3x3 rot_y_matrix = float3x3( + cos_tilt.x, 0.0, sin_tilt.x, + 0.0, 1.0, 0.0, + -sin_tilt.x, 0.0, cos_tilt.x); + static const float3x3 local_to_global = + mul(rot_y_matrix, rot_x_matrix); +/* static const float3x3 global_to_local = + transpose(local_to_global); + const float3x3 local_to_global = float3x3( + cos_tilt.x, sin_tilt.y*sin_tilt.x, cos_tilt.y*sin_tilt.x, + 0.0, cos_tilt.y, sin_tilt.y, + sin_tilt.x, sin_tilt.y*cos_tilt.x, cos_tilt.y*cos_tilt.x); +*/ // This is a pure rotation, so transpose = inverse: + const float3x3 global_to_local = transpose(local_to_global); + // Decompose the matrix into 3 float3's for output: + global_to_local_row0 = float3(global_to_local[0][0], global_to_local[0][1], global_to_local[0][2]);//._m00_m01_m02); + global_to_local_row1 = float3(global_to_local[1][0], global_to_local[1][1], global_to_local[1][2]);//._m10_m11_m12); + global_to_local_row2 = float3(global_to_local[2][0], global_to_local[2][1], global_to_local[2][2]);//._m20_m21_m22); + #else + static const float3x3 global_to_local = geom_global_to_local_static; + static const float3x3 local_to_global = geom_local_to_global_static; + #endif + + // Get an optimal eye position based on geom_view_dist, viewport_aspect, + // and CRT radius/rotation: + #ifdef RUNTIME_GEOMETRY_MODE + const float geom_mode = geom_mode_runtime; + #else + static const float geom_mode = geom_mode_static; + #endif + const float3 eye_pos_global = + get_ideal_global_eye_pos(local_to_global, geom_aspect, geom_mode); + eye_pos_local = mul(global_to_local, eye_pos_global); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/manifest.bml b/shaders/CRT-Royale.shader/manifest.bml new file mode 100644 index 000000000..778c0689c --- /dev/null +++ b/shaders/CRT-Royale.shader/manifest.bml @@ -0,0 +1,214 @@ +input + filter: nearest + +// IMPORTANT: +// Shader passes need to know details about the image in the mask_texture LUT +// files, so set the following constants in user-preset-constants.h accordingly: +// 1.) mask_triads_per_tile = (number of horizontal triads in mask texture LUT's) +// 2.) mask_texture_small_size = (texture size of mask*texture_small LUT's) +// 3.) mask_texture_large_size = (texture size of mask*texture_large LUT's) +// 4.) mask_grille_avg_color = (avg. brightness of mask_grille_texture* LUT's, in [0, 1]) +// 5.) mask_slot_avg_color = (avg. brightness of mask_slot_texture* LUT's, in [0, 1]) +// 6.) mask_shadow_avg_color = (avg. brightness of mask_shadow_texture* LUT's, in [0, 1]) +// Shader passes also need to know certain scales set in this preset, but their +// compilation model doesn't currently allow the preset file to tell them. Make +// sure to set the following constants in user-preset-constants.h accordingly too: +// 1.) bloom_approx_scale_x = scale_x2 +// 2.) mask_resize_viewport_scale = vec2(scale_x6, scale_y5) +// Finally, shader passes need to know the value of geom_max_aspect_ratio used to +// calculate scale_y5 (among other values): +// 1.) geom_max_aspect_ratio = (geom_max_aspect_ratio used to calculate scale_y5) + +// Pass0: Linearize the input based on CRT gamma and bob interlaced fields. +// (Bobbing ensures we can immediately blur without getting artifacts.) +program + filter: nearest + vertex: first-pass-linearize-crt-gamma-bob-fields.vs + fragment: first-pass-linearize-crt-gamma-bob-fields.fs + format: rgba16f + height: 100% + width: 100% + +// Pass1: Resample interlaced (and misconverged) scanlines vertically. +// Separating vertical/horizontal scanline sampling is faster: It lets us +// consider more scanlines while calculating weights for fewer pixels, and +// it reduces our samples from vertical*horizontal to vertical+horizontal. +// This has to come right after ORIG_LINEARIZED, because there's no +// "original_source" scale_type we can use later. +program + filter: linear + vertex: scanlines-vertical-interlacing.vs + fragment: scanlines-vertical-interlacing.fs + height: 400% + width: 100% + format: rgba16f + +// Pass2: Do a small resize blur of ORIG_LINEARIZED at an absolute size, and +// account for convergence offsets. We want to blur a predictable portion of the +// screen to match the phosphor bloom, and absolute scale works best for +// reliable results with a fixed-size bloom. Picking a scale is tricky: +// a.) 400x300 is a good compromise for the "fake-bloom" version: It's low enough +// to blur high-res/interlaced sources but high enough that resampling +// doesn't smear low-res sources too much. +// b.) 320x240 works well for the "real bloom" version: It's 1-1.5% faster, and +// the only noticeable visual difference is a larger halation spread (which +// may be a good thing for people who like to crank it up). +// Note the 4:3 aspect ratio assumes the input has cropped geom_overscan (so it's +// *intended* for an ~4:3 aspect ratio). +program + filter: linear + vertex: bloom-approx.vs + fragment: bloom-approx.fs + format: rgba16f + width: 320 px + height: 240 px + +// Pass3: Vertically blur the input for halation and refractive diffusion. +// Base this on BLOOM_APPROX: This blur should be small and fast, and blurring +// a constant portion of the screen is probably physically correct if the +// viewport resolution is proportional to the simulated CRT size. +program + filter: linear + vertex: blur9fast-vertical.vs + fragment: blur9fast-vertical.fs + format: rgba16f + height: 100% + width: 100% + +// Pass4: Horizontally blur the input for halation and refractive diffusion. +// Note: Using a one-pass 9x9 blur is about 1% slower. +program + filter: linear + vertex: blur9fast-horizontal.vs + fragment: blur9fast-horizontal.fs + format: rgba16f + height: 100% + width: 100% + +// Pass5: Lanczos-resize the phosphor mask vertically. Set the absolute +// scale_x5 == mask_texture_small_size.x (see IMPORTANT above). Larger scales +// will blur, and smaller scales could get nasty. The vertical size must be +// based on the viewport size and calculated carefully to avoid artifacts later. +// First calculate the minimum number of mask tiles we need to draw. +// Since curvature is computed after the scanline masking pass: +// num_resized_mask_tiles = 2.0; +// If curvature were computed in the scanline masking pass (it's not): +// max_mask_texel_border = ~3.0 * (1/3.0 + 4.0*sqrt(2.0) + 0.5 + 1.0); +// max_mask_tile_border = max_mask_texel_border/ +// (min_resized_phosphor_triad_size * mask_triads_per_tile); +// num_resized_mask_tiles = max(2.0, 1.0 + max_mask_tile_border * 2.0); +// At typical values (triad_size >= 2.0, mask_triads_per_tile == 8): +// num_resized_mask_tiles = ~3.8 +// Triad sizes are given in horizontal terms, so we need geom_max_aspect_ratio +// to relate them to vertical resolution. The widest we expect is: +// geom_max_aspect_ratio = 4.0/3.0 // Note: Shader passes need to know this! +// The fewer triads we tile across the screen, the larger each triad will be as a +// fraction of the viewport size, and the larger scale_y5 must be to draw a full +// num_resized_mask_tiles. Therefore, we must decide the smallest number of +// triads we'll guarantee can be displayed on screen. We'll set this according +// to 3-pixel triads at 768p resolution (the lowest anyone's likely to use): +// min_allowed_viewport_triads = 768.0*geom_max_aspect_ratio / 3.0 = 341.333333 +// Now calculate the viewport scale that ensures we can draw resized_mask_tiles: +// min_scale_x = resized_mask_tiles * mask_triads_per_tile / +// min_allowed_viewport_triads +// scale_y5 = geom_max_aspect_ratio * min_scale_x +// # Some code might depend on equal scales: +// scale_x6 = scale_y5 +// Given our default geom_max_aspect_ratio and min_allowed_viewport_triads: +// scale_y5 = 4.0/3.0 * 2.0/(341.33333 / 8.0) = 0.0625 +// IMPORTANT: The scales MUST be calculated in this way. If you wish to change +// geom_max_aspect_ratio, update that constant in user-preset-constants.h! +program + filter: linear + format: rgba16f + width: 64 px + height: 6.25% + vertex: mask-resize-vertical.vs + fragment: mask-resize-vertical.fs + pixmap: textures/TileableLinearApertureGrille15Wide8And5d5SpacingResizeTo64.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearApertureGrille15Wide8And5d5Spacing.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacingResizeTo64.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearShadowMaskEDPResizeTo64.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearShadowMaskEDP.png + filter: linear + wrap: repeat + +// Pass6: Lanczos-resize the phosphor mask horizontally. scale_x6 = scale_y5. +// TODO: Check again if the shaders actually require equal scales. +program + filter: nearest + vertex: mask-resize-horizontal.vs + fragment: mask-resize-horizontal.fs + format: rgba16f + +// Pass7: Resample (misconverged) scanlines horizontally, apply halation, and +// apply the phosphor mask. +program + filter: linear + format: rgba16f + height: 100% + width: 100% + vertex: scanlines-horizontal-apply-mask.vs + fragment: scanlines-horizontal-apply-mask.fs + pixmap: textures/TileableLinearApertureGrille15Wide8And5d5SpacingResizeTo64.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearApertureGrille15Wide8And5d5Spacing.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacingResizeTo64.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearShadowMaskEDPResizeTo64.png + filter: linear + wrap: repeat + pixmap: textures/TileableLinearShadowMaskEDP.png + filter: linear + wrap: repeat + +// Pass 8: Compute a brightpass. This will require reading the final mask. +program + filter: linear + format: rgba16f + vertex: brightpass.vs + fragment: brightpass.fs + +// Pass 9: Blur the brightpass vertically +program + filter: linear + format: rgba16f + vertex: bloom-vertical.vs + fragment: bloom-vertical.fs + +// Pass 10: Blur the brightpass horizontally and combine it with the dimpass: +program + filter: linear + format: rgba16f + height: 100% + width: 100% + vertex: bloom-horizontal-reconstitute.vs + fragment: bloom-horizontal-reconstitute.fs + +// Pass 11: Compute curvature/AA: +program + filter: linear + format: rgba16f + vertex: geometry-aa-last-pass.vs + fragment: geometry-aa-last-pass.fs + +output + filter: nearest \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/mask-resize-horizontal.fs b/shaders/CRT-Royale.shader/mask-resize-horizontal.fs new file mode 100644 index 000000000..8545d587f --- /dev/null +++ b/shaders/CRT-Royale.shader/mask-resize-horizontal.fs @@ -0,0 +1,3208 @@ +#version 150 + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; + +in Vertex { + vec2 vTexCoord; + vec2 src_tex_uv_wrap; + vec2 tile_uv_wrap; + vec2 resize_magnification_scale; + vec2 src_dxdy; + vec2 tile_size_uv; + vec2 input_tiles_per_texture; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + // The input contains one mask tile horizontally and a number vertically. + // Resize the tile horizontally to its final screen size and repeat it + // until drawing at least mask_resize_num_tiles, leaving it unchanged + // vertically. Lanczos-resizing the phosphor mask achieves much sharper + // results than mipmapping, outputting >= mask_resize_num_tiles makes for + // easier tiled sampling later. + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + // Discard unneeded fragments in case our profile allows real branches. + //const float2 tile_uv_wrap = tile_uv_wrap; + if(get_mask_sample_mode() < 0.5 && + max(tile_uv_wrap.x, tile_uv_wrap.y) <= mask_resize_num_tiles) + { + const float src_dx = src_dxdy.x; + const float2 src_tex_uv = frac(src_tex_uv_wrap); + const float3 pixel_color = downsample_horizontal_sinc_tiled(input_texture, + src_tex_uv, texture_size, src_dxdy.x, + resize_magnification_scale.x, tile_size_uv.x); + // The input LUT was linear RGB, and so is our output: + FragColor = float4(pixel_color, 1.0); + } + else + { + discard; + } + #else + discard; + FragColor = float4(1.0,1.0,1.0,1.0); + #endif +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/mask-resize-horizontal.vs b/shaders/CRT-Royale.shader/mask-resize-horizontal.vs new file mode 100644 index 000000000..b64cf9c8c --- /dev/null +++ b/shaders/CRT-Royale.shader/mask-resize-horizontal.vs @@ -0,0 +1,3236 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 src_tex_uv_wrap; + vec2 tile_uv_wrap; + vec2 resize_magnification_scale; + vec2 src_dxdy; + vec2 tile_size_uv; + vec2 input_tiles_per_texture; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; +uniform int phase; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + gl_Position = position; + vTexCoord = texCoord * 1.0001; + float2 tex_uv = vTexCoord.xy; + // First estimate the viewport size (the user will get the wrong number of + // triads if it's wrong and mask_specify_num_triads is 1.0/true). + const float2 estimated_viewport_size = + output_size / mask_resize_viewport_scale; + // Find the final size of our resized phosphor mask tiles. We probably + // estimated the viewport size and MASK_RESIZE output size differently last + // pass, so do not swear they were the same. ;) + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + estimated_viewport_size, output_size, false); + + // We'll render resized tiles until filling the output FBO or meeting a + // limit, so compute [wrapped] tile uv coords based on the output uv coords + // and the number of tiles that will fit in the FBO. + const float2 output_tiles_this_pass = output_size / mask_resize_tile_size; + const float2 output_video_uv = tex_uv * texture_size / video_size; + const float2 tile_uv_wrap = output_video_uv * output_tiles_this_pass; + + // Get the texel size of an input tile and related values: + const float2 input_tile_size = float2(min( + mask_resize_src_lut_size.x, video_size.x), mask_resize_tile_size.y); + tile_size_uv = input_tile_size / texture_size; + input_tiles_per_texture = texture_size / input_tile_size; + + // Derive [wrapped] texture uv coords from [wrapped] tile uv coords and + // the tile size in uv coords, and save frac() for the fragment shader. + src_tex_uv_wrap = tile_uv_wrap * tile_size_uv; + + // Output the values we need, including the magnification scale and step: + //tile_uv_wrap = tile_uv_wrap; + //src_tex_uv_wrap = src_tex_uv_wrap; + resize_magnification_scale = mask_resize_tile_size / input_tile_size; + src_dxdy = float2(1.0/texture_size.x, 0.0); + //tile_size_uv = tile_size_uv; + //input_tiles_per_texture = input_tiles_per_texture; +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/mask-resize-vertical.fs b/shaders/CRT-Royale.shader/mask-resize-vertical.fs new file mode 100644 index 000000000..16e8090ef --- /dev/null +++ b/shaders/CRT-Royale.shader/mask-resize-vertical.fs @@ -0,0 +1,3248 @@ +#version 150 + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; +uniform sampler2D pixmap[]; +uniform int phase; + +in Vertex { + vec2 vTexCoord; + vec2 src_tex_uv_wrap; + vec2 resize_magnification_scale; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define mask_grille_texture_small pixmap[0] +#define mask_slot_texture_small pixmap[2] +#define mask_shadow_texture_small pixmap[4] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + // Resize the input phosphor mask tile to the final vertical size it will + // appear on screen. Keep 1x horizontal size if possible (IN.output_size + // >= mask_resize_src_lut_size), and otherwise linearly sample horizontally + // to fit exactly one tile. Lanczos-resizing the phosphor mask achieves + // much sharper results than mipmapping, and vertically resizing first + // minimizes the total number of taps required. We output a number of + // resized tiles >= mask_resize_num_tiles for easier tiled sampling later. + //const float2 src_tex_uv_wrap = src_tex_uv_wrap; + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + // Discard unneeded fragments in case our profile allows real branches. + const float2 tile_uv_wrap = src_tex_uv_wrap; + if(get_mask_sample_mode() < 0.5 && + tile_uv_wrap.y <= mask_resize_num_tiles) + { + static const float src_dy = 1.0/mask_resize_src_lut_size.y; + const float2 src_tex_uv = frac(src_tex_uv_wrap); + float3 pixel_color; + // If mask_type is static, this branch will be resolved statically. + #ifdef PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + if(mask_type < 0.5) + { + pixel_color = downsample_vertical_sinc_tiled( + mask_grille_texture_large, src_tex_uv, mask_resize_src_lut_size, + src_dy, resize_magnification_scale.y, 1.0); + } + else if(mask_type < 1.5) + { + pixel_color = downsample_vertical_sinc_tiled( + mask_slot_texture_large, src_tex_uv, mask_resize_src_lut_size, + src_dy, resize_magnification_scale.y, 1.0); + } + else + { + pixel_color = downsample_vertical_sinc_tiled( + mask_shadow_texture_large, src_tex_uv, mask_resize_src_lut_size, + src_dy, resize_magnification_scale.y, 1.0); + } + #else + if(mask_type < 0.5) + { + pixel_color = downsample_vertical_sinc_tiled( + mask_grille_texture_small, src_tex_uv, mask_resize_src_lut_size, + src_dy, resize_magnification_scale.y, 1.0); + } + else if(mask_type < 1.5) + { + pixel_color = downsample_vertical_sinc_tiled( + mask_slot_texture_small, src_tex_uv, mask_resize_src_lut_size, + src_dy, resize_magnification_scale.y, 1.0); + } + else + { + pixel_color = downsample_vertical_sinc_tiled( + mask_shadow_texture_small, src_tex_uv, mask_resize_src_lut_size, + src_dy, resize_magnification_scale.y, 1.0); + } + #endif + // The input LUT was linear RGB, and so is our output: + FragColor = float4(pixel_color, 1.0); + } + else + { + discard; + } + #else + discard; + FragColor = float4(1.0, 1.0, 1.0, 1.0); + #endif +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/mask-resize-vertical.vs b/shaders/CRT-Royale.shader/mask-resize-vertical.vs new file mode 100644 index 000000000..2dac429b4 --- /dev/null +++ b/shaders/CRT-Royale.shader/mask-resize-vertical.vs @@ -0,0 +1,3212 @@ +#version 150 + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 src_tex_uv_wrap; + vec2 resize_magnification_scale; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; +uniform int phase; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + gl_Position = position; + vTexCoord = texCoord; + float2 tex_uv = vTexCoord.xy; + // First estimate the viewport size (the user will get the wrong number of + // triads if it's wrong and mask_specify_num_triads is 1.0/true). + const float viewport_y = output_size.y / mask_resize_viewport_scale.y; + const float aspect_ratio = geom_aspect_ratio_x / geom_aspect_ratio_y; + const float2 estimated_viewport_size = + float2(viewport_y * aspect_ratio, viewport_y); + // Estimate the output size of MASK_RESIZE (the next pass). The estimated + // x component shouldn't matter, because we're not using the x result, and + // we're not swearing it's correct (if we did, the x result would influence + // the y result to maintain the tile aspect ratio). + const float2 estimated_mask_resize_output_size = + float2(output_size.y * aspect_ratio, output_size.y); + // Find the final intended [y] size of our resized phosphor mask tiles, + // then the tile size for the current pass (resize y only): + float2 mask_resize_tile_size = get_resized_mask_tile_size( + estimated_viewport_size, estimated_mask_resize_output_size, false); + float2 pass_output_tile_size = float2(min( + mask_resize_src_lut_size.x, output_size.x), mask_resize_tile_size.y); + + // We'll render resized tiles until filling the output FBO or meeting a + // limit, so compute [wrapped] tile uv coords based on the output uv coords + // and the number of tiles that will fit in the FBO. + const float2 output_tiles_this_pass = output_size / pass_output_tile_size; + const float2 output_video_uv = tex_uv * texture_size / video_size; + const float2 tile_uv_wrap = output_video_uv * output_tiles_this_pass; + + // The input LUT is just a single mask tile, so texture uv coords are the + // same as tile uv coords (save frac() for the fragment shader). The + // magnification scale is also straightforward: + src_tex_uv_wrap = tile_uv_wrap; + resize_magnification_scale = + pass_output_tile_size / mask_resize_src_lut_size; +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/scanlines-horizontal-apply-mask.fs b/shaders/CRT-Royale.shader/scanlines-horizontal-apply-mask.fs new file mode 100644 index 000000000..a987afbbf --- /dev/null +++ b/shaders/CRT-Royale.shader/scanlines-horizontal-apply-mask.fs @@ -0,0 +1,10845 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; +uniform int phase; +uniform sampler2D pixmap[]; + +in Vertex { + vec2 vTexCoord; + vec2 video_uv; + vec2 scanline_tex_uv; + vec2 blur3x3_tex_uv; + vec2 halation_tex_uv; + vec2 scanline_texture_size_inv; + vec4 mask_tile_start_uv_and_size; + vec2 mask_tiles_per_screen; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +#define VERTICAL_SCANLINEStexture source[5] +#define VERTICAL_SCANLINEStexture_size sourceSize[5].xy +#define VERTICAL_SCANLINESvideo_size sourceSize[5].xy +#define BLOOM_APPROXtexture source[4] +#define BLOOM_APPROXtexture_size sourceSize[4].xy +#define BLOOM_APPROXvideo_size sourceSize[4].xy +#define HALATION_BLURtexture source[2] +#define HALATION_BLURtexture_size sourceSize[2].xy +#define HALATION_BLURvideo_size sourceSize[2].xy +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #define MASK_RESIZEtexture source[0] +#else + #define MASK_RESIZEtexture source[0] +#endif +#define MASK_RESIZEtexture_size sourceSize[0] +#define MASK_RESIZEvideo_size sourceSize[0] + +#define input_texture source[0] +#define mask_grille_texture_large pixmap[1] +#define mask_slot_texture_large pixmap[3] +#define mask_shadow_texture_large pixmap[5] + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +// VERTEX INCLUDES // + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +//#include "../../../../include/gamma-management.h" +// already got it + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 tex2Dtiled_mask_linearize(const sampler2D tex, + const float2 tex_uv) +{ + // If we're manually tiling a texture, anisotropic filtering can get + // confused. One workaround is to just select the lowest mip level: + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + // TODO: Use tex2Dlod_linearize with a calculated mip level. + return tex2Dlod_linearize(tex, float4(tex_uv, 0.0, 0.0)); + #else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + return tex2Dbias_linearize(tex, float4(tex_uv, 0.0, -16.0)); + #else + return tex2D_linearize(tex, tex_uv); + #endif + #endif + #else + return tex2D_linearize(tex, tex_uv); + #endif +} + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +// END VERTEX INCLUDES // + +////////////////////////////// FRAGMENT INCLUDES ////////////////////////////// + +//#include "bloom-functions.h" + +//////////////////////////// BEGIN BLOOM-FUNCTIONS /////////////////////////// + +#ifndef BLOOM_FUNCTIONS_H +#define BLOOM_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These utility functions and constants help several passes determine the +// size and center texel weight of the phosphor bloom in a uniform manner. + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// We need to calculate the correct blur sigma using some .cgp constants: +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/blur-functions.h" + +//////////////////////////// BEGIN BLUR-FUNCTIONS /////////////////////////// + +#ifndef BLUR_FUNCTIONS_H +#define BLUR_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides reusable one-pass and separable (two-pass) blurs. +// Requires: All blurs share these requirements (dxdy requirement is split): +// 1.) All requirements of gamma-management.h must be satisfied! +// 2.) filter_linearN must == "true" in your .cgp preset unless +// you're using tex2DblurNresize at 1x scale. +// 3.) mipmap_inputN must == "true" in your .cgp preset if +// output_size < video_size. +// 4.) output_size == video_size / pow(2, M), where M is some +// positive integer. tex2Dblur*resize can resize arbitrarily +// (and the blur will be done after resizing), but arbitrary +// resizes "fail" with other blurs due to the way they mix +// static weights with bilinear sample exploitation. +// 5.) In general, dxdy should contain the uv pixel spacing: +// dxdy = (video_size/output_size)/texture_size +// 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), +// zero out the dxdy component in the unblurred dimension: +// dxdy = float2(dxdy.x, 0.0) or float2(0.0, dxdy.y) +// Many blurs share these requirements: +// 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, +// or they will blur more in the lower-scaled dimension. +// 2.) One-pass shared sample blurs require ddx(), ddy(), and +// tex2Dlod() to be supported by the current Cg profile, and +// the drivers must support high-quality derivatives. +// 3.) One-pass shared sample blurs require: +// tex_uv.w == log2(video_size/output_size).y; +// Non-wrapper blurs share this requirement: +// 1.) sigma is the intended standard deviation of the blur +// Wrapper blurs share this requirement, which is automatically +// met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): +// 1.) blurN_std_dev must be global static const float values +// specifying standard deviations for Nx blurs in units +// of destination pixels +// Optional: 1.) The including file (or an earlier included file) may +// optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace +// default standard deviations with those matching a binomial +// distribution. (See below for details/properties.) +// 2.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_BLUR_STD_DEVS and override: +// static const float blur3_std_dev +// static const float blur4_std_dev +// static const float blur5_std_dev +// static const float blur6_std_dev +// static const float blur7_std_dev +// static const float blur8_std_dev +// static const float blur9_std_dev +// static const float blur10_std_dev +// static const float blur11_std_dev +// static const float blur12_std_dev +// static const float blur17_std_dev +// static const float blur25_std_dev +// static const float blur31_std_dev +// static const float blur43_std_dev +// 3.) The including file (or an earlier included file) may +// optionally #define OVERRIDE_ERROR_BLURRING and override: +// static const float error_blurring +// This tuning value helps mitigate weighting errors from one- +// pass shared-sample blurs sharing bilinear samples between +// fragments. Values closer to 0.0 have "correct" blurriness +// but allow more artifacts, and values closer to 1.0 blur away +// artifacts by sampling closer to halfway between texels. +// UPDATE 6/21/14: The above static constants may now be overridden +// by non-static uniform constants. This permits exposing blur +// standard deviations as runtime GUI shader parameters. However, +// using them keeps weights from being statically computed, and the +// speed hit depends on the blur: On my machine, uniforms kill over +// 53% of the framerate with tex2Dblur12x12shared, but they only +// drop the framerate by about 18% with tex2Dblur11fast. +// Quality and Performance Comparisons: +// For the purposes of the following discussion, "no sRGB" means +// GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. +// 1.) tex2DblurNfast is always faster than tex2DblurNresize. +// 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize +// well, because they're the only ones that don't exploit bilinear samples. +// This also means they're the only functions which can be truly gamma- +// correct without linear (or sRGB FBO) input, but only at 1x scale. +// 3.) One-pass shared sample blurs only have a speed advantage without sRGB. +// They also have some inaccuracies due to their shared-[bilinear-]sample +// design, which grow increasingly bothersome for smaller blurs and higher- +// frequency source images (relative to their resolution). I had high +// hopes for them, but their most realistic use case is limited to quickly +// reblurring an already blurred input at full resolution. Otherwise: +// a.) If you're blurring a low-resolution source, you want a better blur. +// b.) If you're blurring a lower mipmap, you want a better blur. +// c.) If you're blurring a high-resolution, high-frequency source, you +// want a better blur. +// 4.) The one-pass blurs without shared samples grow slower for larger blurs, +// but they're competitive with separable blurs at 5x5 and smaller, and +// even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. +// Here are some framerates from a GeForce 8800GTS. The first pass resizes to +// viewport size (4x in this test) and linearizes for sRGB codepaths, and the +// remaining passes perform 6 full blurs. Mipmapped tests are performed at the +// same scale, so they just measure the cost of mipmapping each FBO (only every +// other FBO is mipmapped for separable blurs, to mimic realistic usage). +// Mipmap Neither sRGB+Mipmap sRGB Function +// 76.0 92.3 131.3 193.7 tex2Dblur3fast +// 63.2 74.4 122.4 175.5 tex2Dblur3resize +// 93.7 121.2 159.3 263.2 tex2Dblur3x3 +// 59.7 68.7 115.4 162.1 tex2Dblur3x3resize +// 63.2 74.4 122.4 175.5 tex2Dblur5fast +// 49.3 54.8 100.0 132.7 tex2Dblur5resize +// 59.7 68.7 115.4 162.1 tex2Dblur5x5 +// 64.9 77.2 99.1 137.2 tex2Dblur6x6shared +// 55.8 63.7 110.4 151.8 tex2Dblur7fast +// 39.8 43.9 83.9 105.8 tex2Dblur7resize +// 40.0 44.2 83.2 104.9 tex2Dblur7x7 +// 56.4 65.5 71.9 87.9 tex2Dblur8x8shared +// 49.3 55.1 99.9 132.5 tex2Dblur9fast +// 33.3 36.2 72.4 88.0 tex2Dblur9resize +// 27.8 29.7 61.3 72.2 tex2Dblur9x9 +// 37.2 41.1 52.6 60.2 tex2Dblur10x10shared +// 44.4 49.5 91.3 117.8 tex2Dblur11fast +// 28.8 30.8 63.6 75.4 tex2Dblur11resize +// 33.6 36.5 40.9 45.5 tex2Dblur12x12shared +// TODO: Fill in benchmarks for new untested blurs. +// tex2Dblur17fast +// tex2Dblur25fast +// tex2Dblur31fast +// tex2Dblur43fast +// tex2Dblur3x3resize + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +// Set static standard deviations, but allow users to override them with their +// own constants (even non-static uniforms if they're okay with the speed hit): +#ifndef OVERRIDE_BLUR_STD_DEVS + // blurN_std_dev values are specified in terms of dxdy strides. + #ifdef USE_BINOMIAL_BLUR_STD_DEVS + // By request, we can define standard deviations corresponding to a + // binomial distribution with p = 0.5 (related to Pascal's triangle). + // This distribution works such that blurring multiple times should + // have the same result as a single larger blur. These values are + // larger than default for blurs up to 6x and smaller thereafter. + static const float blur3_std_dev = 0.84931640625; + static const float blur4_std_dev = 0.84931640625; + static const float blur5_std_dev = 1.0595703125; + static const float blur6_std_dev = 1.06591796875; + static const float blur7_std_dev = 1.17041015625; + static const float blur8_std_dev = 1.1720703125; + static const float blur9_std_dev = 1.2259765625; + static const float blur10_std_dev = 1.21982421875; + static const float blur11_std_dev = 1.25361328125; + static const float blur12_std_dev = 1.2423828125; + static const float blur17_std_dev = 1.27783203125; + static const float blur25_std_dev = 1.2810546875; + static const float blur31_std_dev = 1.28125; + static const float blur43_std_dev = 1.28125; + #else + // The defaults are the largest values that keep the largest unused + // blur term on each side <= 1.0/256.0. (We could get away with more + // or be more conservative, but this compromise is pretty reasonable.) + static const float blur3_std_dev = 0.62666015625; + static const float blur4_std_dev = 0.66171875; + static const float blur5_std_dev = 0.9845703125; + static const float blur6_std_dev = 1.02626953125; + static const float blur7_std_dev = 1.36103515625; + static const float blur8_std_dev = 1.4080078125; + static const float blur9_std_dev = 1.7533203125; + static const float blur10_std_dev = 1.80478515625; + static const float blur11_std_dev = 2.15986328125; + static const float blur12_std_dev = 2.215234375; + static const float blur17_std_dev = 3.45535583496; + static const float blur25_std_dev = 5.3409576416; + static const float blur31_std_dev = 6.86488037109; + static const float blur43_std_dev = 10.1852050781; + #endif // USE_BINOMIAL_BLUR_STD_DEVS +#endif // OVERRIDE_BLUR_STD_DEVS + +#ifndef OVERRIDE_ERROR_BLURRING + // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing + // in shared-sample blurs but increase blurring and feature shifting. + static const float error_blurring = 0.5; +#endif + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +// gamma-management.h relies on pass-specific settings to guide its behavior: +// FIRST_PASS, LAST_PASS, GAMMA_ENCODE_EVERY_FBO, etc. See it for details. +//#include "gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//#include "quad-pixel-communication.h" + +/////////////////////// BEGIN QUAD-PIXEL-COMMUNICATION ////////////////////// + +#ifndef QUAD_PIXEL_COMMUNICATION_H +#define QUAD_PIXEL_COMMUNICATION_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey* +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DISCLAIMER ///////////////////////////////// + +// *This code was inspired by "Shader Amortization using Pixel Quad Message +// Passing" by Eric Penner, published in GPU Pro 2, Chapter VI.2. My intent +// is not to plagiarize his fundamentally similar code and assert my own +// copyright, but the algorithmic helper functions require so little code that +// implementations can't vary by much except bugfixes and conventions. I just +// wanted to license my own particular code here to avoid ambiguity and make it +// clear that as far as I'm concerned, people can do as they please with it. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// Given screen pixel numbers, derive a "quad vector" describing a fragment's +// position in its 2x2 pixel quad. Given that vector, obtain the values of any +// variable at neighboring fragments. +// Requires: Using this file in general requires: +// 1.) ddx() and ddy() are present in the current Cg profile. +// 2.) The GPU driver is using fine/high-quality derivatives. +// Functions will give incorrect results if this is not true, +// so a test function is included. + + +///////////////////// QUAD-PIXEL COMMUNICATION PRIMITIVES //////////////////// + +float4 get_quad_vector_naive(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Two measures of the current fragment's output pixel number + // in the range ([0, output_size.x), [0, output_size.y)): + // 1.) output_pixel_num_wrt_uvxy.xy increase with uv coords. + // 2.) output_pixel_num_wrt_uvxy.zw increase with screen xy. + // Returns: Two measures of the fragment's position in its 2x2 quad: + // 1.) The .xy components are its 2x2 placement with respect to + // uv direction (the origin (0, 0) is at the top-left): + // top-left = (-1.0, -1.0) top-right = ( 1.0, -1.0) + // bottom-left = (-1.0, 1.0) bottom-right = ( 1.0, 1.0) + // You need this to arrange/weight shared texture samples. + // 2.) The .zw components are its 2x2 placement with respect to + // screen xy direction (position); the origin varies. + // quad_gather needs this measure to work correctly. + // Note: quad_vector.zw = quad_vector.xy * float2( + // ddx(output_pixel_num_wrt_uvxy.x), + // ddy(output_pixel_num_wrt_uvxy.y)); + // Caveats: This function assumes the GPU driver always starts 2x2 pixel + // quads at even pixel numbers. This assumption can be wrong + // for odd output resolutions (nondeterministically so). + float4 pixel_odd = frac(output_pixel_num_wrt_uvxy * 0.5) * 2.0; + float4 quad_vector = pixel_odd * 2.0 - float4(1.0); + return quad_vector; +} + +float4 get_quad_vector(float4 output_pixel_num_wrt_uvxy) +{ + // Requires: Same as get_quad_vector_naive() (see that first). + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + float4 quad_vector_guess = + get_quad_vector_naive(output_pixel_num_wrt_uvxy); + // If quad_vector_guess.zw doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_guess.z), + ddy(quad_vector_guess.w)); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +float4 get_quad_vector(float2 output_pixel_num_wrt_uv) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) output_pixel_num_wrt_uv must increase with uv coords and + // measure the current fragment's output pixel number in: + // ([0, output_size.x), [0, output_size.y)) + // Returns: Same as get_quad_vector_naive() (see that first), but it's + // correct even if the 2x2 pixel quad starts at an odd pixel, + // which can occur at odd resolutions. + // Caveats: This function requires less information than the version + // taking a float4, but it's potentially slower. + // Do screen coords increase with or against uv? Get the direction + // with respect to (uv.x, uv.y) for (screen.x, screen.y) in {-1, 1}. + float2 screen_uv_mirror = float2(ddx(output_pixel_num_wrt_uv.x), + ddy(output_pixel_num_wrt_uv.y)); + float2 pixel_odd_wrt_uv = frac(output_pixel_num_wrt_uv * 0.5) * 2.0; + float2 quad_vector_uv_guess = (pixel_odd_wrt_uv - float2(0.5)) * 2.0; + float2 quad_vector_screen_guess = quad_vector_uv_guess * screen_uv_mirror; + // If quad_vector_screen_guess doesn't increase with screen xy, we know + // the 2x2 pixel quad starts at an odd pixel: + float2 odd_start_mirror = 0.5 * float2(ddx(quad_vector_screen_guess.x), + ddy(quad_vector_screen_guess.y)); + float4 quad_vector_guess = float4( + quad_vector_uv_guess, quad_vector_screen_guess); + return quad_vector_guess * odd_start_mirror.xyxy; +} + +void quad_gather(float4 quad_vector, float4 curr, + out float4 adjx, out float4 adjy, out float4 diag) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) The GPU driver is using fine/high-quality derivatives. + // 3.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 4.) curr is any vector you wish to get neighboring values of. + // Returns: Values of an input vector (curr) at neighboring fragments + // adjacent x, adjacent y, and diagonal (via out parameters). + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float3 curr, + out float3 adjx, out float3 adjy, out float3 diag) +{ + // Float3 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +void quad_gather(float4 quad_vector, float2 curr, + out float2 adjx, out float2 adjy, out float2 diag) +{ + // Float2 version + adjx = curr - ddx(curr) * quad_vector.z; + adjy = curr - ddy(curr) * quad_vector.w; + diag = adjx - ddy(adjx) * quad_vector.w; +} + +float4 quad_gather(float4 quad_vector, float curr) +{ + // Float version: + // Returns: return.x == current + // return.y == adjacent x + // return.z == adjacent y + // return.w == diagonal + float4 all = float4(curr); + all.y = all.x - ddx(all.x) * quad_vector.z; + all.zw = all.xy - ddy(all.xy) * quad_vector.w; + return all; +} + +float4 quad_gather_sum(float4 quad_vector, float4 curr) +{ + // Requires: Same as quad_gather() + // Returns: Sum of an input vector (curr) at all fragments in a quad. + float4 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float3 quad_gather_sum(float4 quad_vector, float3 curr) +{ + // Float3 version: + float3 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float2 quad_gather_sum(float4 quad_vector, float2 curr) +{ + // Float2 version: + float2 adjx, adjy, diag; + quad_gather(quad_vector, curr, adjx, adjy, diag); + return (curr + adjx + adjy + diag); +} + +float quad_gather_sum(float4 quad_vector, float curr) +{ + // Float version: + float4 all_values = quad_gather(quad_vector, curr); + return (all_values.x + all_values.y + all_values.z + all_values.w); +} + +bool fine_derivatives_working(float4 quad_vector, float4 curr) +{ + // Requires: 1.) ddx() and ddy() are present in the current Cg profile. + // 2.) quad_vector describes the current fragment's location in + // its 2x2 pixel quad using get_quad_vector()'s conventions. + // 3.) curr must be a test vector with non-constant derivatives + // (its value should change nonlinearly across fragments). + // Returns: true if fine/hybrid/high-quality derivatives are used, or + // false if coarse derivatives are used or inconclusive + // Usage: Test whether quad-pixel communication is working! + // Method: We can confirm fine derivatives are used if the following + // holds (ever, for any value at any fragment): + // (ddy(curr) != ddy(adjx)) or (ddx(curr) != ddx(adjy)) + // The more values we test (e.g. test a float4 two ways), the + // easier it is to demonstrate fine derivatives are working. + // TODO: Check for floating point exact comparison issues! + float4 ddx_curr = ddx(curr); + float4 ddy_curr = ddy(curr); + float4 adjx = curr - ddx_curr * quad_vector.z; + float4 adjy = curr - ddy_curr * quad_vector.w; + bool ddy_different = any(bool4(ddy_curr.x != ddy(adjx).x, ddy_curr.y != ddy(adjx).y, ddy_curr.z != ddy(adjx).z, ddy_curr.w != ddy(adjx).w)); + bool ddx_different = any(bool4(ddx_curr.x != ddx(adjy).x, ddx_curr.y != ddx(adjy).y, ddx_curr.z != ddx(adjy).z, ddx_curr.w != ddx(adjy).w)); + return any(bool2(ddy_different, ddx_different)); +} + +bool fine_derivatives_working_fast(float4 quad_vector, float curr) +{ + // Requires: Same as fine_derivatives_working() + // Returns: Same as fine_derivatives_working() + // Usage: This is faster than fine_derivatives_working() but more + // likely to return false negatives, so it's less useful for + // offline testing/debugging. It's also useless as the basis + // for dynamic runtime branching as of May 2014: Derivatives + // (and quad-pixel communication) are currently disallowed in + // branches. However, future GPU's may allow you to use them + // in dynamic branches if you promise the branch condition + // evaluates the same for every fragment in the quad (and/or if + // the driver enforces that promise by making a single fragment + // control branch decisions). If that ever happens, this + // version may become a more economical choice. + float ddx_curr = ddx(curr); + float ddy_curr = ddy(curr); + float adjx = curr - ddx_curr * quad_vector.z; + return (ddy_curr != ddy(adjx)); +} + +#endif // QUAD_PIXEL_COMMUNICATION_H + +//////////////////////// END QUAD-PIXEL-COMMUNICATION /////////////////////// + +//#include "special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 uv2_to_uv4(float2 tex_uv) +{ + // Make a float2 uv offset safe for adding to float4 tex2Dlod coords: + return float4(tex_uv, 0.0, 0.0); +} + +// Make a length squared helper macro (for usage with static constants): +#define LENGTH_SQ(vec) (dot(vec, vec)) + +inline float get_fast_gaussian_weight_sum_inv(const float sigma) +{ + // We can use the Gaussian integral to calculate the asymptotic weight for + // the center pixel. Since the unnormalized center pixel weight is 1.0, + // the normalized weight is the same as the weight sum inverse. Given a + // large enough blur (9+), the asymptotic weight sum is close and faster: + // center_weight = 0.5 * + // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) + // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): + // However, we can get even faster results with curve-fitting. These are + // also closer than the asymptotic results, because they were constructed + // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from + // (0, blurN_std_dev), so the results for smaller sigmas are biased toward + // smaller blurs. The max error is 0.0031793913. + // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. + //static const float temp = 0.5/sqrt(2.0); + //return erf(temp/sigma); + return min(exp(exp(0.348348412457428/ + (sigma - 0.0860587260734721))), 0.399334576340352/sigma); +} + + +//////////////////// ARBITRARILY RESIZABLE SEPARABLE BLURS /////////////////// + +float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 11x Gaussian blurred texture lookup using a 11-tap blur. + // It may be mipmapped depending on settings and dxdy. + // Calculate Gaussian blur kernel weights and a normalization factor for + // distances of 0-4, ignoring constant factors (since we're normalizing). + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Statically normalize weights, sum weighted samples, and return. Blurs are + // currently optimized for dynamic weights. + float3 sum = float3(0.0,0.0,0.0); + sum += w5 * tex2D_linearize(tex, tex_uv - 5.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + sum += w5 * tex2D_linearize(tex, tex_uv + 5.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 9x Gaussian blurred texture lookup using a 9-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w4 * tex2D_linearize(tex, tex_uv - 4.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + sum += w4 * tex2D_linearize(tex, tex_uv + 4.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 7x Gaussian blurred texture lookup using a 7-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w3 * tex2D_linearize(tex, tex_uv - 3.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + sum += w3 * tex2D_linearize(tex, tex_uv + 3.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 5x Gaussian blurred texture lookup using a 5-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w2 * tex2D_linearize(tex, tex_uv - 2.0 * dxdy).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + sum += w2 * tex2D_linearize(tex, tex_uv + 2.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 1D 3x Gaussian blurred texture lookup using a 3-tap blur. + // It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * tex2D_linearize(tex, tex_uv - 1.0 * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1 * tex2D_linearize(tex, tex_uv + 1.0 * dxdy).rgb; + return sum * weight_sum_inv; +} + + +/////////////////////////// FAST SEPARABLE BLURS /////////////////////////// + +float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: 1.) Global requirements must be met (see file description). + // 2.) filter_linearN must = "true" in your .cgp file. + // 3.) For gamma-correct bilinear filtering, global + // gamma_aware_bilinear == true (from gamma-management.h) + // Returns: A 1D 11x Gaussian blurred texture lookup using 6 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w45 = w4 + w5; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + const float w45_ratio = w5/w45; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w45 * tex2D_linearize(tex, tex_uv - (4.0 + w45_ratio) * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + sum += w45 * tex2D_linearize(tex, tex_uv + (4.0 + w45_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 9x Gaussian blurred texture lookup using 1 nearest + // neighbor and 4 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w34 = w3 + w4; + const float w12_ratio = w2/w12; + const float w34_ratio = w4/w34; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w34 * tex2D_linearize(tex, tex_uv - (3.0 + w34_ratio) * dxdy).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + sum += w34 * tex2D_linearize(tex, tex_uv + (3.0 + w34_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 7x Gaussian blurred texture lookup using 4 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3)); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w23 = w2 + w3; + const float w01_ratio = w1/w01; + const float w23_ratio = w3/w23; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w23 * tex2D_linearize(tex, tex_uv - (2.0 + w23_ratio) * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb; + sum += w01 * tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb; + sum += w23 * tex2D_linearize(tex, tex_uv + (2.0 + w23_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 5x Gaussian blurred texture lookup using 1 nearest + // neighbor and 2 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2)); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w12 = w1 + w2; + const float w12_ratio = w2/w12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w12 * tex2D_linearize(tex, tex_uv - (1.0 + w12_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w12 * tex2D_linearize(tex, tex_uv + (1.0 + w12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 3x Gaussian blurred texture lookup using 2 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float weight_sum_inv = 1.0 / (w0 + 2.0 * w1); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w01 = w0 * 0.5 + w1; + const float w01_ratio = w1/w01; + // Weights for all samples are the same, so just average them: + return 0.5 * ( + tex2D_linearize(tex, tex_uv - w01_ratio * dxdy).rgb + + tex2D_linearize(tex, tex_uv + w01_ratio * dxdy).rgb); +} + + +//////////////////////////// HUGE SEPARABLE BLURS //////////////////////////// + +// Huge separable blurs come only in "fast" versions. +float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 43x Gaussian blurred texture lookup using 22 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + + // w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w16_17 = w16 + w17; + const float w18_19 = w18 + w19; + const float w20_21 = w20 + w21; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + const float w16_17_ratio = w17/w16_17; + const float w18_19_ratio = w19/w18_19; + const float w20_21_ratio = w21/w20_21; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w20_21 * tex2D_linearize(tex, tex_uv - (20.0 + w20_21_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv - (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv - (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w16_17 * tex2D_linearize(tex, tex_uv + (16.0 + w16_17_ratio) * dxdy).rgb; + sum += w18_19 * tex2D_linearize(tex, tex_uv + (18.0 + w18_19_ratio) * dxdy).rgb; + sum += w20_21 * tex2D_linearize(tex, tex_uv + (20.0 + w20_21_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 31x Gaussian blurred texture lookup using 16 linear + // taps. It may be mipmapped depending on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + //const float weight_sum_inv = 1.0 / + // (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + + // w9 + w10 + w11 + w12 + w13 + w14 + w15)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + // The center texel (with weight w0) is used twice, so halve its weight. + const float w0_1 = w0 * 0.5 + w1; + const float w2_3 = w2 + w3; + const float w4_5 = w4 + w5; + const float w6_7 = w6 + w7; + const float w8_9 = w8 + w9; + const float w10_11 = w10 + w11; + const float w12_13 = w12 + w13; + const float w14_15 = w14 + w15; + const float w0_1_ratio = w1/w0_1; + const float w2_3_ratio = w3/w2_3; + const float w4_5_ratio = w5/w4_5; + const float w6_7_ratio = w7/w6_7; + const float w8_9_ratio = w9/w8_9; + const float w10_11_ratio = w11/w10_11; + const float w12_13_ratio = w13/w12_13; + const float w14_15_ratio = w15/w14_15; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w14_15 * tex2D_linearize(tex, tex_uv - (14.0 + w14_15_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv - (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv - (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv - (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv - (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv - (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv - (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv - w0_1_ratio * dxdy).rgb; + sum += w0_1 * tex2D_linearize(tex, tex_uv + w0_1_ratio * dxdy).rgb; + sum += w2_3 * tex2D_linearize(tex, tex_uv + (2.0 + w2_3_ratio) * dxdy).rgb; + sum += w4_5 * tex2D_linearize(tex, tex_uv + (4.0 + w4_5_ratio) * dxdy).rgb; + sum += w6_7 * tex2D_linearize(tex, tex_uv + (6.0 + w6_7_ratio) * dxdy).rgb; + sum += w8_9 * tex2D_linearize(tex, tex_uv + (8.0 + w8_9_ratio) * dxdy).rgb; + sum += w10_11 * tex2D_linearize(tex, tex_uv + (10.0 + w10_11_ratio) * dxdy).rgb; + sum += w12_13 * tex2D_linearize(tex, tex_uv + (12.0 + w12_13_ratio) * dxdy).rgb; + sum += w14_15 * tex2D_linearize(tex, tex_uv + (14.0 + w14_15_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 25x Gaussian blurred texture lookup using 1 nearest + // neighbor and 12 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w9_10 = w9 + w10; + const float w11_12 = w11 + w12; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + const float w9_10_ratio = w10/w9_10; + const float w11_12_ratio = w12/w11_12; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w11_12 * tex2D_linearize(tex, tex_uv - (11.0 + w11_12_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv - (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w9_10 * tex2D_linearize(tex, tex_uv + (9.0 + w9_10_ratio) * dxdy).rgb; + sum += w11_12 * tex2D_linearize(tex, tex_uv + (11.0 + w11_12_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + +float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Same as tex2Dblur11() + // Returns: A 1D 17x Gaussian blurred texture lookup using 1 nearest + // neighbor and 8 linear taps. It may be mipmapped depending + // on settings and dxdy. + // First get the texel weights and normalization factor as above. + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + //const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + // w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + const float weight_sum_inv = get_fast_gaussian_weight_sum_inv(sigma); + // Calculate combined weights and linear sample ratios between texel pairs. + const float w1_2 = w1 + w2; + const float w3_4 = w3 + w4; + const float w5_6 = w5 + w6; + const float w7_8 = w7 + w8; + const float w1_2_ratio = w2/w1_2; + const float w3_4_ratio = w4/w3_4; + const float w5_6_ratio = w6/w5_6; + const float w7_8_ratio = w8/w7_8; + // Statically normalize weights, sum weighted samples, and return: + float3 sum = float3(0.0,0.0,0.0); + sum += w7_8 * tex2D_linearize(tex, tex_uv - (7.0 + w7_8_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv - (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv - (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv - (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w0 * tex2D_linearize(tex, tex_uv).rgb; + sum += w1_2 * tex2D_linearize(tex, tex_uv + (1.0 + w1_2_ratio) * dxdy).rgb; + sum += w3_4 * tex2D_linearize(tex, tex_uv + (3.0 + w3_4_ratio) * dxdy).rgb; + sum += w5_6 * tex2D_linearize(tex, tex_uv + (5.0 + w5_6_ratio) * dxdy).rgb; + sum += w7_8 * tex2D_linearize(tex, tex_uv + (7.0 + w7_8_ratio) * dxdy).rgb; + return sum * weight_sum_inv; +} + + +//////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// + +float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Requires: Global requirements must be met (see file description). + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the + // resized input. + // Description: + // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize + // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. + const float denom_inv = 0.5/(sigma*sigma); + // Load each sample. We need all 3x3 samples. Quad-pixel communication + // won't help either: This should perform like tex2Dblur5x5, but sharing a + // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). + const float2 sample4_uv = tex_uv; + const float2 dx = float2(dxdy.x, 0.0); + const float2 dy = float2(0.0, dxdy.y); + const float2 sample1_uv = sample4_uv - dy; + const float2 sample7_uv = sample4_uv + dy; + const float3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; + const float3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; + const float3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; + const float3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; + const float3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; + const float3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; + const float3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; + const float3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; + const float3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; + // Statically compute Gaussian sample weights: + const float w4 = 1.0; + const float w1_3_5_7 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w0_2_6_8 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); + // Weight and sum the samples: + const float3 sum = w4 * sample4 + + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); + return sum * weight_sum_inv; +} + + +//////////////////////////// FASTER ONE-PASS BLURS /////////////////////////// + +float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 9x9 Gaussian blurred mipmapped texture lookup composed of + // 5x5 carefully selected bilinear samples. + // Description: + // Perform a 1-pass 9x9 blur with 5x5 bilinear samples. Adjust the + // bilinear sample location to reflect the true Gaussian weights for each + // underlying texel. The following diagram illustrates the relative + // locations of bilinear samples. Each sample with the same number has the + // same weight (notice the symmetry). The letters a, b, c, d distinguish + // quadrants, and the letters U, D, L, R, C (up, down, left, right, center) + // distinguish 1D directions along the line containing the pixel center: + // 6a 5a 2U 5b 6b + // 4a 3a 1U 3b 4b + // 2L 1L 0C 1R 2R + // 4c 3c 1D 3d 4d + // 6c 5c 2D 5d 6d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2, 2x1, 1x2, or 1x1 texel block: + // 6a4 6a3 5a4 5a3 2U2 5b3 5b4 6b3 6b4 + // 6a2 6a1 5a2 5a1 2U1 5b1 5b2 6b1 6b2 + // 4a4 4a3 3a4 3a3 1U2 3b3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3a1 1U1 3b1 3b2 4b1 4b2 + // 2L2 2L1 1L2 1L1 0C1 1R1 1R2 2R1 2R2 + // 4c2 4c1 3c2 3c1 1D1 3d1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3c3 1D2 3d3 3d4 4d3 4d4 + // 6c2 6c1 5c2 5c1 2D1 5d1 5d2 6d1 6d2 + // 6c4 6c3 5c4 5c3 2D2 5d3 5d4 6d3 6d4 + // Note there is only one C texel and only two texels for each U, D, L, or + // R sample. The center sample is effectively a nearest neighbor sample, + // and the U/D/L/R samples use 1D linear filtering. All other texels are + // read with bilinear samples somewhere within their 2x2 texel blocks. + + // COMPUTE TEXTURE COORDS: + // Statically compute sampling offsets within each 2x2 texel block, based + // on 1D sampling ratios between texels [1, 2] and [3, 4] texels away from + // the center, and reuse them independently for both dimensions. Compute + // these offsets based on the relative 1D Gaussian weights of the texels + // in question. (w1off means "Gaussian weight for the texel 1.0 texels + // away from the pixel center," etc.). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float w4off = exp(-16.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + const float texel3to4ratio = w4off/(w3off + w4off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2R_texel_offset = float2(3.0, 0.0) + float2(texel3to4ratio, 0.0); + const float2 sample3d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + const float2 sample4d_texel_offset = float2(3.0, 1.0) + float2(texel3to4ratio, texel1to2ratio); + const float2 sample5d_texel_offset = float2(1.0, 3.0) + float2(texel1to2ratio, texel3to4ratio); + const float2 sample6d_texel_offset = float2(3.0, 3.0) + float2(texel3to4ratio, texel3to4ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2R1 = w3off; + const float w2R2 = w4off; + const float w3d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w3d2_3d3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w3d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d1_5d1 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d2_5d3 = exp(-LENGTH_SQ(float2(4.0, 1.0)) * denom_inv); + const float w4d3_5d2 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4_5d4 = exp(-LENGTH_SQ(float2(4.0, 2.0)) * denom_inv); + const float w6d1 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + const float w6d2_6d3 = exp(-LENGTH_SQ(float2(4.0, 3.0)) * denom_inv); + const float w6d4 = exp(-LENGTH_SQ(float2(4.0, 4.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2R1 + w2R2; + const float w3 = w3d1 + 2.0 * w3d2_3d3 + w3d4; + const float w4 = w4d1_5d1 + w4d2_5d3 + w4d3_5d2 + w4d4_5d4; + const float w5 = w4; + const float w6 = w6d1 + 2.0 * w6d2_6d3 + w6d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(w0 + 4.0 * (w1 + w2 + w3 + w4 + w5 + w6)); + + // LOAD TEXTURE SAMPLES: + // Load all 25 samples (1 nearest, 8 linear, 16 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + // Sampling order doesn't seem to affect performance, so just be clear: + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2R = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset).rgb; + const float3 sample2D = tex2D_linearize(tex, tex_uv + dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample2L = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset).rgb; + const float3 sample2U = tex2D_linearize(tex, tex_uv - dxdy * sample2R_texel_offset.yx).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample5d = tex2D_linearize(tex, tex_uv + dxdy * sample5d_texel_offset).rgb; + const float3 sample5c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample5d_texel_offset).rgb; + const float3 sample5b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample5d_texel_offset).rgb; + const float3 sample5a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample5d_texel_offset).rgb; + const float3 sample6d = tex2D_linearize(tex, tex_uv + dxdy * sample6d_texel_offset).rgb; + const float3 sample6c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample6d_texel_offset).rgb; + const float3 sample6b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample6d_texel_offset).rgb; + const float3 sample6a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample6d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2R + sample2D + sample2L + sample2U); + sum += w3 * (sample3d + sample3c + sample3b + sample3a); + sum += w4 * (sample4d + sample4c + sample4b + sample4a); + sum += w5 * (sample5d + sample5c + sample5b + sample5a); + sum += w6 * (sample6d + sample6c + sample6b + sample6a); + return sum * weight_sum_inv; +} + +float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 7x7 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 7x7 Gaussian blurred mipmapped texture lookup composed of + // 4x4 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 4a 3a 3b 4b + // 2a 1a 1b 2b + // 2c 1c 1d 2d + // 4c 3c 3d 4d + // The texel layout is as follows. Note that samples 3a/3b, 1a/1b, 1c/1d, + // and 3c/3d share a vertical column of texels, and samples 2a/2c, 1a/1c, + // 1b/1d, and 2b/2d share a horizontal row of texels (all sample1's share + // the center texel): + // 4a4 4a3 3a4 3ab3 3b4 4b3 4b4 + // 4a2 4a1 3a2 3ab1 3b2 4b1 4b2 + // 2a4 2a3 1a4 1ab3 1b4 2b3 2b4 + // 2ac2 2ac1 1ac2 1* 1bd2 2bd1 2bd2 + // 2c4 2c3 1c4 1cd3 1d4 2d3 2d4 + // 4c2 4c1 3c2 3cd1 3d2 4d1 4d2 + // 4c4 4c3 3c4 3cd3 3d4 4d3 4d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float w3off = exp(-9.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + const float texel2to3ratio = w3off/(w2off + w3off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample1d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + const float2 sample2d_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample3d_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4d_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1abcd = 1.0; + const float w1bd2_1cd3 = exp(-LENGTH_SQ(float2(1.0, 0.0)) * denom_inv); + const float w2bd1_3cd1 = exp(-LENGTH_SQ(float2(2.0, 0.0)) * denom_inv); + const float w2bd2_3cd2 = exp(-LENGTH_SQ(float2(3.0, 0.0)) * denom_inv); + const float w1d4 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d3_3d2 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4_3d4 = exp(-LENGTH_SQ(float2(3.0, 1.0)) * denom_inv); + const float w4d1 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + const float w4d2_4d3 = exp(-LENGTH_SQ(float2(3.0, 2.0)) * denom_inv); + const float w4d4 = exp(-LENGTH_SQ(float2(3.0, 3.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights. + // Split weights for shared texels between samples sharing them: + const float w1 = w1abcd * 0.25 + w1bd2_1cd3 + w1d4; + const float w2_3 = (w2bd1_3cd1 + w2bd2_3cd2) * 0.5 + w2d3_3d2 + w2d4_3d4; + const float w4 = w4d1 + 2.0 * w4d2_4d3 + w4d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = + 1.0/(4.0 * (w1 + 2.0 * w2_3 + w4)); + + // LOAD TEXTURE SAMPLES: + // Load all 16 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample1a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample1d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + const float3 sample3a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample3d_texel_offset).rgb; + const float3 sample4a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample4d_texel_offset).rgb; + const float3 sample1b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample1d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample3b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample3d_texel_offset).rgb; + const float3 sample4b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample4d_texel_offset).rgb; + const float3 sample1c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample1d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample3c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample3d_texel_offset).rgb; + const float3 sample4c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample4d_texel_offset).rgb; + const float3 sample1d = tex2D_linearize(tex, tex_uv + dxdy * sample1d_texel_offset).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample3d = tex2D_linearize(tex, tex_uv + dxdy * sample3d_texel_offset).rgb; + const float3 sample4d = tex2D_linearize(tex, tex_uv + dxdy * sample4d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = float3(0.0,0.0,0.0); + sum += w1 * (sample1a + sample1b + sample1c + sample1d); + sum += w2_3 * (sample2a + sample2b + sample2c + sample2d); + sum += w2_3 * (sample3a + sample3b + sample3c + sample3d); + sum += w4 * (sample4a + sample4b + sample4c + sample4d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 5x5 blur with 3x3 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 5x5 Gaussian blurred mipmapped texture lookup composed of + // 3x3 carefully selected bilinear samples. + // Description: + // First see the description for tex2Dblur9x9(). This blur uses the same + // concept and sample/texel locations except on a smaller scale. Samples: + // 2a 1U 2b + // 1L 0C 1R + // 2c 1D 2d + // Texels: + // 2a4 2a3 1U2 2b3 2b4 + // 2a2 2a1 1U1 2b1 2b2 + // 1L2 1L1 0C1 1R1 1R2 + // 2c2 2c1 1D1 2d1 2d2 + // 2c4 2c3 1D2 2d3 2d4 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w1off = exp(-1.0 * denom_inv); + const float w2off = exp(-4.0 * denom_inv); + const float texel1to2ratio = w2off/(w1off + w2off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including x-axis-aligned: + const float2 sample1R_texel_offset = float2(1.0, 0.0) + float2(texel1to2ratio, 0.0); + const float2 sample2d_texel_offset = float2(1.0, 1.0) + float2(texel1to2ratio, texel1to2ratio); + + // CALCULATE KERNEL WEIGHTS FOR ALL SAMPLES: + // Statically compute Gaussian texel weights for the bottom-right quadrant. + // Read underscores as "and." + const float w1R1 = w1off; + const float w1R2 = w2off; + const float w2d1 = exp(-LENGTH_SQ(float2(1.0, 1.0)) * denom_inv); + const float w2d2_3 = exp(-LENGTH_SQ(float2(2.0, 1.0)) * denom_inv); + const float w2d4 = exp(-LENGTH_SQ(float2(2.0, 2.0)) * denom_inv); + // Statically add texel weights in each sample to get sample weights: + const float w0 = 1.0; + const float w1 = w1R1 + w1R2; + const float w2 = w2d1 + 2.0 * w2d2_3 + w2d4; + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0 + 4.0 * (w1 + w2)); + + // LOAD TEXTURE SAMPLES: + // Load all 9 samples (1 nearest, 4 linear, 4 bilinear) using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0C = tex2D_linearize(tex, tex_uv).rgb; + const float3 sample1R = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset).rgb; + const float3 sample1D = tex2D_linearize(tex, tex_uv + dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample1L = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset).rgb; + const float3 sample1U = tex2D_linearize(tex, tex_uv - dxdy * sample1R_texel_offset.yx).rgb; + const float3 sample2d = tex2D_linearize(tex, tex_uv + dxdy * sample2d_texel_offset).rgb; + const float3 sample2c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample2d_texel_offset).rgb; + const float3 sample2b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample2d_texel_offset).rgb; + const float3 sample2a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample2d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Statically normalize weights (so total = 1.0), and sum weighted samples. + float3 sum = w0 * sample0C; + sum += w1 * (sample1R + sample1D + sample1L + sample1U); + sum += w2 * (sample2a + sample2b + sample2c + sample2d); + return sum * weight_sum_inv; +} + +float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // Perform a 1-pass 3x3 blur with 5x5 bilinear samples. + // Requires: Same as tex2Dblur9() + // Returns: A 3x3 Gaussian blurred mipmapped texture lookup composed of + // 2x2 carefully selected bilinear samples. + // Description: + // First see the descriptions for tex2Dblur9x9() and tex2Dblur7(). This + // blur mixes concepts from both. The sample layout is as follows: + // 0a 0b + // 0c 0d + // The texel layout is as follows. Note that samples 0a/0b and 0c/0d share + // a vertical column of texels, and samples 0a/0c and 0b/0d share a + // horizontal row of texels (all samples share the center texel): + // 0a3 0ab2 0b3 + // 0ac1 0*0 0bd1 + // 0c3 0cd2 0d3 + + // COMPUTE TEXTURE COORDS: + // Statically compute bilinear sampling offsets (details in tex2Dblur9x9). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w1off = exp(-1.0 * denom_inv); + const float texel0to1ratio = w1off/(w0off * 0.5 + w1off); + // Statically compute texel offsets from the fragment center to each + // bilinear sample in the bottom-right quadrant, including axis-aligned: + const float2 sample0d_texel_offset = float2(texel0to1ratio, texel0to1ratio); + + // LOAD TEXTURE SAMPLES: + // Load all 4 samples using symmetry: + const float2 mirror_x = float2(-1.0, 1.0); + const float2 mirror_y = float2(1.0, -1.0); + const float2 mirror_xy = float2(-1.0, -1.0); + const float2 dxdy_mirror_x = dxdy * mirror_x; + const float2 dxdy_mirror_y = dxdy * mirror_y; + const float2 dxdy_mirror_xy = dxdy * mirror_xy; + const float3 sample0a = tex2D_linearize(tex, tex_uv + dxdy_mirror_xy * sample0d_texel_offset).rgb; + const float3 sample0b = tex2D_linearize(tex, tex_uv + dxdy_mirror_y * sample0d_texel_offset).rgb; + const float3 sample0c = tex2D_linearize(tex, tex_uv + dxdy_mirror_x * sample0d_texel_offset).rgb; + const float3 sample0d = tex2D_linearize(tex, tex_uv + dxdy * sample0d_texel_offset).rgb; + + // SUM WEIGHTED SAMPLES: + // Weights for all samples are the same, so just average them: + return 0.25 * (sample0a + sample0b + sample0c + sample0d); +} + + +////////////////// LINEAR ONE-PASS BLURS WITH SHARED SAMPLES ///////////////// + +float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: 1.) Same as tex2Dblur9() + // 2.) ddx() and ddy() are present in the current Cg profile. + // 3.) The GPU driver is using fine/high-quality derivatives. + // 4.) quad_vector *correctly* describes the current fragment's + // location in its pixel quad, by the conventions noted in + // get_quad_vector[_naive]. + // 5.) tex_uv.w = log2(video_size/output_size).y + // 6.) tex2Dlod() is present in the current Cg profile. + // Optional: Tune artifacts vs. excessive blurriness with the global + // float error_blurring. + // Returns: A blurred texture lookup using a "virtual" 12x12 Gaussian + // blur (a 6x6 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // Perform a 1-pass blur with shared texture lookups across a pixel quad. + // We'll get neighboring samples with high-quality ddx/ddy derivatives, as + // in GPU Pro 2, Chapter VI.2, "Shader Amortization using Pixel Quad + // Message Passing" by Eric Penner. + // + // Our "virtual" 12x12 blur will be comprised of ((6 - 1)^2)/4 + 3 = 12 + // bilinear samples, where bilinear sampling positions are computed from + // the relative Gaussian weights of the 4 surrounding texels. The catch is + // that the appropriate texel weights and sample coords differ for each + // fragment, but we're reusing most of the same samples across a quad of + // destination fragments. (We do use unique coords for the four nearest + // samples at each fragment.) Mixing bilinear filtering and sample-sharing + // therefore introduces some error into the weights, and this can get nasty + // when the source image is small or high-frequency. Computing bilinear + // ratios based on weights at the sample field center results in sharpening + // and ringing artifacts, but we can move samples closer to halfway between + // texels to try blurring away the error (which can move features around by + // a texel or so). Tune this with the global float "error_blurring". + // + // The pixel quad's sample field covers 12x12 texels, accessed through 6x6 + // bilinear (2x2 texel) taps. Each fragment depends on a window of 10x10 + // texels (5x5 bilinear taps), and each fragment is responsible for loading + // a 6x6 texel quadrant as a 3x3 block of bilinear taps, plus 3 more taps + // to use unique bilinear coords for sample0* for each fragment. This + // diagram illustrates the relative locations of bilinear samples 1-9 for + // each quadrant a, b, c, d (note samples will not be equally spaced): + // 8a 7a 6a 6b 7b 8b + // 5a 4a 3a 3b 4b 5b + // 2a 1a 0a 0b 1b 2b + // 2c 1c 0c 0d 1d 2d + // 5c 4c 3c 3d 4d 5d + // 8c 7c 6c 6d 7d 8d + // The following diagram illustrates the underlying equally spaced texels, + // named after the sample that accesses them and subnamed by their location + // within their 2x2 texel block: + // 8a3 8a2 7a3 7a2 6a3 6a2 6b2 6b3 7b2 7b3 8b2 8b3 + // 8a1 8a0 7a1 7a0 6a1 6a0 6b0 6b1 7b0 7b1 8b0 8b1 + // 5a3 5a2 4a3 4a2 3a3 3a2 3b2 3b3 4b2 4b3 5b2 5b3 + // 5a1 5a0 4a1 4a0 3a1 3a0 3b0 3b1 4b0 4b1 5b0 5b1 + // 2a3 2a2 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 2b2 2b3 + // 2a1 2a0 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 2b0 2b1 + // 2c1 2c0 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 2d0 2d1 + // 2c3 2c2 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 2d2 2d3 + // 5c1 5c0 4c1 4c0 3c1 3c0 3d0 3d1 4d0 4d1 5d0 5d1 + // 5c3 5c2 4c3 4c2 3c3 3c2 3d2 3d3 4d2 4d3 5d2 5d3 + // 8c1 8c0 7c1 7c0 6c1 6c0 6d0 6d1 7d0 7d1 8d0 8d1 + // 8c3 8c2 7c3 7c2 6c3 6c2 6d2 6d3 7d2 7d3 8d2 8d3 + // With this symmetric arrangement, we don't have to know which absolute + // quadrant a sample lies in to assign kernel weights; it's enough to know + // the sample number and the relative quadrant of the sample (relative to + // the current quadrant): + // {current, adjacent x, adjacent y, diagonal} + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute sampling offsets within each 2x2 texel block, based + // on appropriate 1D Gaussian sampling ratio between texels [0, 1], [2, 3], + // and [4, 5] away from the fragment, and reuse them independently for both + // dimensions. Use the sample field center as the estimated destination, + // but nudge the result closer to halfway between texels to blur error. + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // based on the sum of their 4 underlying texel weights. Assume a same- + // resolution blur, so each symmetrically named sample weight will compute + // the same at every fragment in the pixel quad: We can therefore compute + // texel weights based only on the bottom-right quadrant (fragment at 0d0). + // Too avoid too much boilerplate code, use a macro to get all 4 texel + // weights for a bilinear sample based on the offset of its top-left texel: + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w8diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -6.0); + const float w7diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -6.0); + const float w6diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -6.0); + const float w6adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -6.0); + const float w7adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -6.0); + const float w8adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -6.0); + const float w5diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -4.0); + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-6.0, -2.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 0.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w5adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 2.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w8adjx = GET_TEXEL_QUAD_WEIGHTS(-6.0, 4.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w5 = float4(w5curr, w5adjx, w5adjy, w5diag); + const float4 w6 = float4(w6curr, w6adjx, w6adjy, w6diag); + const float4 w7 = float4(w7curr, w7adjx, w7adjy, w7diag); + const float4 w8 = float4(w8curr, w8adjx, w8adjy, w8diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + float3 sample8adjx, sample8adjy, sample8diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + quad_gather(quad_vector, sample8curr, sample8adjx, sample8adjy, sample8diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + sum += mul(w5, float4x3(sample5curr, sample5adjx, sample5adjy, sample5diag)); + sum += mul(w6, float4x3(sample6curr, sample6adjx, sample6adjy, sample6diag)); + sum += mul(w7, float4x3(sample7curr, sample7adjx, sample7adjy, sample7diag)); + sum += mul(w8, float4x3(sample8curr, sample8adjx, sample8adjy, sample8diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 10x10 Gaussian + // blur (a 5x5 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 25 of the 36 samples taken across the pixel quad (to cover a + // 5x5 sample area, or 10x10 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 11 omitted samples + // are always the "same:" + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float w4_5off = exp(-(4.5*4.5) * denom_inv); + const float w5_5off = exp(-(5.5*5.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + const float texel4to5ratio = lerp(w5_5off/(w4_5off + w5_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(4.0, 0.0) + float2(texel4to5ratio, texel0to1ratio); + const float2 sample3_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample4_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + const float2 sample5_texel_offset = float2(4.0, 2.0) + float2(texel4to5ratio, texel2to3ratio); + const float2 sample6_texel_offset = float2(0.0, 4.0) + float2(texel0to1ratio, texel4to5ratio); + const float2 sample7_texel_offset = float2(2.0, 4.0) + float2(texel2to3ratio, texel4to5ratio); + const float2 sample8_texel_offset = float2(4.0, 4.0) + float2(texel4to5ratio, texel4to5ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 25 of the 36 sample weights. Skip the following weights: + // 8adjx, 2adjx, 5adjx, + // 6adjy, 7adjy, 8adjy, + // 2diag, 5diag, 6diag, 7diag, 8diag + const float w4diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w4adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w5adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(4.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 0.0); + const float w4adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w4curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + const float w5curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 2.0); + const float w7adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 4.0); + const float w6adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 4.0); + const float w6curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 4.0); + const float w7curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 4.0); + const float w8curr = GET_TEXEL_QUAD_WEIGHTS(4.0, 4.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w4curr + w5curr + w6curr + w7curr + w8curr + + w0adjx + w1adjx + w3adjx + w4adjx + w6adjx + w7adjx + + w0adjy + w1adjy + w2adjy + w3adjy + w4adjy + w5adjy + + w0diag + w1diag + w3diag + w4diag); + // Statically pack most weights for runtime. Note the mixed packing: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + const float4 w4 = float4(w4curr, w4adjx, w4adjy, w4diag); + const float4 w2and5 = float4(w2curr, w2adjy, w5curr, w5adjy); + const float4 w6and7 = float4(w6curr, w6adjx, w7curr, w7adjx); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + const float3 sample4curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample4_texel_offset)).rgb; + const float3 sample5curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample5_texel_offset)).rgb; + const float3 sample6curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample6_texel_offset)).rgb; + const float3 sample7curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample7_texel_offset)).rgb; + const float3 sample8curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample8_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad in order of need: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + float3 sample4adjx, sample4adjy, sample4diag; + float3 sample5adjx, sample5adjy, sample5diag; + float3 sample6adjx, sample6adjy, sample6diag; + float3 sample7adjx, sample7adjy, sample7diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + quad_gather(quad_vector, sample4curr, sample4adjx, sample4adjy, sample4diag); + quad_gather(quad_vector, sample5curr, sample5adjx, sample5adjy, sample5diag); + quad_gather(quad_vector, sample6curr, sample6adjx, sample6adjy, sample6diag); + quad_gather(quad_vector, sample7curr, sample7adjx, sample7adjy, sample7diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result. First do the simple ones: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + sum += mul(w4, float4x3(sample4curr, sample4adjx, sample4adjy, sample4diag)); + // Now do the mixed-sample ones: + sum += mul(w2and5, float4x3(sample2curr, sample2adjy, sample5curr, sample5adjy)); + sum += mul(w6and7, float4x3(sample6curr, sample6adjx, sample7curr, sample7adjx)); + sum += w8curr * sample8curr; + // Normalize the sum (so the weights add to 1.0) and return: + return sum * weight_sum_inv; +} + +float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 8x8 Gaussian + // blur (a 4x4 blur of carefully selected bilinear samples) + // of the given mip level. There will be subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur12x12shared(). This function + // shares the same concept and a similar sample placement, except each + // quadrant contains 4x4 texels and 2x2 samples instead of 6x6 and 3x3 + // respectively. There could be a total of 16 samples, 4 of which each + // fragment is responsible for, but each fragment loads 0a/0b/0c/0d with + // its own offset to reduce shared sample artifacts, bringing the sample + // count for each fragment to 7. Sample placement: + // 3a 2a 2b 3b + // 1a 0a 0b 1b + // 1c 0c 0d 1d + // 3c 2c 2d 3d + // Texel placement: + // 3a3 3a2 2a3 2a2 2b2 2b3 3b2 3b3 + // 3a1 3a0 2a1 2a0 2b0 2b1 3b0 3b1 + // 1a3 1a2 0a3 0a2 0b2 0b3 1b2 1b3 + // 1a1 1a0 0a1 0a0 0b0 0b1 1b0 1b1 + // 1c1 1c0 0c1 0c0 0d0 0d1 1d0 1d1 + // 1c3 1c2 0c3 0c2 0d2 0d3 1d2 1d3 + // 3c1 3c0 2c1 2c0 2d0 2d1 3d0 4d1 + // 3c3 3c2 2c3 2c2 2d2 2d3 3d2 4d3 + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + const float w3diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -4.0); + const float w2diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -4.0); + const float w2adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -4.0); + const float w3adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -4.0); + const float w1diag = GET_TEXEL_QUAD_WEIGHTS(-4.0, -2.0); + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w1adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 0.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w3adjx = GET_TEXEL_QUAD_WEIGHTS(-4.0, 2.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Statically pack weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + const float4 w1 = float4(w1curr, w1adjx, w1adjy, w1diag); + const float4 w2 = float4(w2curr, w2adjx, w2adjy, w2diag); + const float4 w3 = float4(w3curr, w3adjx, w3adjy, w3diag); + // Get the weight sum inverse (normalization factor): + const float4 weight_sum4 = w0 + w1 + w2 + w3; + const float2 weight_sum2 = weight_sum4.xy + weight_sum4.zw; + const float weight_sum = weight_sum2.x + weight_sum2.y; + const float weight_sum_inv = 1.0/(weight_sum); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + float3 sample3adjx, sample3adjy, sample3diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + quad_gather(quad_vector, sample3curr, sample3adjx, sample3adjy, sample3diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += mul(w1, float4x3(sample1curr, sample1adjx, sample1adjy, sample1diag)); + sum += mul(w2, float4x3(sample2curr, sample2adjx, sample2adjy, sample2diag)); + sum += mul(w3, float4x3(sample3curr, sample3adjx, sample3adjy, sample3diag)); + return sum * weight_sum_inv; +} + +float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector, + const float sigma) +{ + // Perform a 1-pass mipmapped blur with shared samples across a pixel quad. + // Requires: Same as tex2Dblur12x12shared() + // Returns: A blurred texture lookup using a "virtual" 6x6 Gaussian + // blur (a 3x3 blur of carefully selected bilinear samples) + // of the given mip level. There will be some inaccuracies,subtle inaccuracies, + // especially for small or high-frequency detailed sources. + // Description: + // First see the description for tex2Dblur8x8shared(). This + // function shares the same concept and sample placement, but each fragment + // only uses 9 of the 16 samples taken across the pixel quad (to cover a + // 3x3 sample area, or 6x6 texel area), and it uses a lower standard + // deviation to compensate. Thanks to symmetry, the 7 omitted samples + // are always the "same:" + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + + // COMPUTE COORDS FOR TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Statically compute bilinear sampling offsets (details in tex2Dblur12x12shared). + const float denom_inv = 0.5/(sigma*sigma); + const float w0off = 1.0; + const float w0_5off = exp(-(0.5*0.5) * denom_inv); + const float w1off = exp(-(1.0*1.0) * denom_inv); + const float w1_5off = exp(-(1.5*1.5) * denom_inv); + const float w2off = exp(-(2.0*2.0) * denom_inv); + const float w2_5off = exp(-(2.5*2.5) * denom_inv); + const float w3_5off = exp(-(3.5*3.5) * denom_inv); + const float texel0to1ratio = lerp(w1_5off/(w0_5off + w1_5off), 0.5, error_blurring); + const float texel2to3ratio = lerp(w3_5off/(w2_5off + w3_5off), 0.5, error_blurring); + // We don't share sample0*, so use the nearest destination fragment: + const float texel0to1ratio_nearest = w1off/(w0off + w1off); + const float texel1to2ratio_nearest = w2off/(w1off + w2off); + // Statically compute texel offsets from the bottom-right fragment to each + // bilinear sample in the bottom-right quadrant: + const float2 sample0curr_texel_offset = float2(0.0, 0.0) + float2(texel0to1ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjx_texel_offset = float2(-1.0, 0.0) + float2(-texel1to2ratio_nearest, texel0to1ratio_nearest); + const float2 sample0adjy_texel_offset = float2(0.0, -1.0) + float2(texel0to1ratio_nearest, -texel1to2ratio_nearest); + const float2 sample0diag_texel_offset = float2(-1.0, -1.0) + float2(-texel1to2ratio_nearest, -texel1to2ratio_nearest); + const float2 sample1_texel_offset = float2(2.0, 0.0) + float2(texel2to3ratio, texel0to1ratio); + const float2 sample2_texel_offset = float2(0.0, 2.0) + float2(texel0to1ratio, texel2to3ratio); + const float2 sample3_texel_offset = float2(2.0, 2.0) + float2(texel2to3ratio, texel2to3ratio); + + // CALCULATE KERNEL WEIGHTS: + // Statically compute bilinear sample weights at each destination fragment + // from the sum of their 4 texel weights (details in tex2Dblur12x12shared). + #define GET_TEXEL_QUAD_WEIGHTS(xoff, yoff) \ + (exp(-LENGTH_SQ(float2(xoff, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff, yoff + 1.0)) * denom_inv) + \ + exp(-LENGTH_SQ(float2(xoff + 1.0, yoff + 1.0)) * denom_inv)) + // We only need 9 of the 16 sample weights. Skip the following weights: + // 1adjx, 3adjx + // 2adjy, 3adjy + // 1diag, 2diag, 3diag + const float w0diag = GET_TEXEL_QUAD_WEIGHTS(-2.0, -2.0); + const float w0adjy = GET_TEXEL_QUAD_WEIGHTS(0.0, -2.0); + const float w1adjy = GET_TEXEL_QUAD_WEIGHTS(2.0, -2.0); + const float w0adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 0.0); + const float w0curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 0.0); + const float w1curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 0.0); + const float w2adjx = GET_TEXEL_QUAD_WEIGHTS(-2.0, 2.0); + const float w2curr = GET_TEXEL_QUAD_WEIGHTS(0.0, 2.0); + const float w3curr = GET_TEXEL_QUAD_WEIGHTS(2.0, 2.0); + #undef GET_TEXEL_QUAD_WEIGHTS + // Get the weight sum inverse (normalization factor): + const float weight_sum_inv = 1.0/(w0curr + w1curr + w2curr + w3curr + + w0adjx + w2adjx + w0adjy + w1adjy + w0diag); + // Statically pack some weights for runtime: + const float4 w0 = float4(w0curr, w0adjx, w0adjy, w0diag); + + // LOAD TEXTURE SAMPLES THIS FRAGMENT IS RESPONSIBLE FOR: + // Get a uv vector from texel 0q0 of this quadrant to texel 0q3: + const float2 dxdy_curr = dxdy * quad_vector.xy; + // Load bilinear samples for the current quadrant (for this fragment): + const float3 sample0curr = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0curr_texel_offset).rgb; + const float3 sample0adjx = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjx_texel_offset).rgb; + const float3 sample0adjy = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0adjy_texel_offset).rgb; + const float3 sample0diag = tex2D_linearize(tex, tex_uv.xy + dxdy_curr * sample0diag_texel_offset).rgb; + const float3 sample1curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample1_texel_offset)).rgb; + const float3 sample2curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample2_texel_offset)).rgb; + const float3 sample3curr = tex2Dlod_linearize(tex, tex_uv + uv2_to_uv4(dxdy_curr * sample3_texel_offset)).rgb; + + // GATHER NEIGHBORING SAMPLES AND SUM WEIGHTED SAMPLES: + // Fetch the samples from other fragments in the 2x2 quad: + float3 sample1adjx, sample1adjy, sample1diag; + float3 sample2adjx, sample2adjy, sample2diag; + quad_gather(quad_vector, sample1curr, sample1adjx, sample1adjy, sample1diag); + quad_gather(quad_vector, sample2curr, sample2adjx, sample2adjy, sample2diag); + // Statically normalize weights (so total = 1.0), and sum weighted samples. + // Fill each row of a matrix with an rgb sample and pre-multiply by the + // weights to obtain a weighted result for sample1*, and handle the rest + // of the weights more directly/verbosely: + float3 sum = float3(0.0,0.0,0.0); + sum += mul(w0, float4x3(sample0curr, sample0adjx, sample0adjy, sample0diag)); + sum += w1curr * sample1curr + w1adjy * sample1adjy + w2curr * sample2curr + + w2adjx * sample2adjx + w3curr * sample3curr; + return sum * weight_sum_inv; +} + + +/////////////////////// MAX OPTIMAL SIGMA BLUR WRAPPERS ////////////////////// + +// The following blurs are static wrappers around the dynamic blurs above. +// HOPEFULLY, the compiler will be smart enough to do constant-folding. + +// Resizable separable blurs: +inline float3 tex2Dblur11resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11resize(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9resize(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7resize(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5resize(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// Fast separable blurs: +inline float3 tex2Dblur11fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur11fast(tex, tex_uv, dxdy, blur11_std_dev); +} +inline float3 tex2Dblur9fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9fast(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7fast(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5fast(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3fast(tex, tex_uv, dxdy, blur3_std_dev); +} +// Huge, "fast" separable blurs: +inline float3 tex2Dblur43fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur43fast(tex, tex_uv, dxdy, blur43_std_dev); +} +inline float3 tex2Dblur31fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur31fast(tex, tex_uv, dxdy, blur31_std_dev); +} +inline float3 tex2Dblur25fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur25fast(tex, tex_uv, dxdy, blur25_std_dev); +} +inline float3 tex2Dblur17fast(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur17fast(tex, tex_uv, dxdy, blur17_std_dev); +} +// Resizable one-pass blurs: +inline float3 tex2Dblur3x3resize(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3resize(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" one-pass blurs: +inline float3 tex2Dblur9x9(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur9x9(tex, tex_uv, dxdy, blur9_std_dev); +} +inline float3 tex2Dblur7x7(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur7x7(tex, tex_uv, dxdy, blur7_std_dev); +} +inline float3 tex2Dblur5x5(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur5x5(tex, tex_uv, dxdy, blur5_std_dev); +} +inline float3 tex2Dblur3x3(const sampler2D tex, const float2 tex_uv, + const float2 dxdy) +{ + return tex2Dblur3x3(tex, tex_uv, dxdy, blur3_std_dev); +} +// "Fast" shared-sample one-pass blurs: +inline float3 tex2Dblur12x12shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur12x12shared(tex, tex_uv, dxdy, quad_vector, blur12_std_dev); +} +inline float3 tex2Dblur10x10shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur10x10shared(tex, tex_uv, dxdy, quad_vector, blur10_std_dev); +} +inline float3 tex2Dblur8x8shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur8x8shared(tex, tex_uv, dxdy, quad_vector, blur8_std_dev); +} +inline float3 tex2Dblur6x6shared(const sampler2D tex, + const float4 tex_uv, const float2 dxdy, const float4 quad_vector) +{ + return tex2Dblur6x6shared(tex, tex_uv, dxdy, quad_vector, blur6_std_dev); +} + + +#endif // BLUR_FUNCTIONS_H + +//////////////////////////// END BLUR-FUNCTIONS /////////////////////////// + +/////////////////////////////// BLOOM CONSTANTS ////////////////////////////// + +// Compute constants with manual inlines of the functions below: +static const float bloom_diff_thresh = 1.0/256.0; + + + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float get_min_sigma_to_blur_triad(const float triad_size, + const float thresh) +{ + // Requires: 1.) triad_size is the final phosphor triad size in pixels + // 2.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum sigma that will fully blur a phosphor + // triad on the screen to an even color, within thresh. + // This closed-form function was found by curve-fitting data. + // Estimate: max error = ~0.086036, mean sq. error = ~0.0013387: + return -0.05168 + 0.6113*triad_size - + 1.122*triad_size*sqrt(0.000416 + thresh); + // Estimate: max error = ~0.16486, mean sq. error = ~0.0041041: + //return 0.5985*triad_size - triad_size*sqrt(thresh) +} + +inline float get_absolute_scale_blur_sigma(const float thresh) +{ + // Requires: 1.) min_expected_triads must be a global float. The number + // of horizontal phosphor triads in the final image must be + // >= min_allowed_viewport_triads.x for realistic results. + // 2.) bloom_approx_scale_x must be a global float equal to the + // absolute horizontal scale of BLOOM_APPROX. + // 3.) bloom_approx_scale_x/min_allowed_viewport_triads.x + // should be <= 1.1658025090 to keep the final result < + // 0.62666015625 (the largest sigma ensuring the largest + // unused texel weight stays < 1.0/256.0 for a 3x3 blur). + // 4.) thresh is the max desired pixel difference in the + // blurred triad (e.g. 1.0/256.0). + // Returns: Return the minimum Gaussian sigma that will blur the pass + // output as much as it would have taken to blur away + // bloom_approx_scale_x horizontal phosphor triads. + // Description: + // BLOOM_APPROX should look like a downscaled phosphor blur. Ideally, we'd + // use the same blur sigma as the actual phosphor bloom and scale it down + // to the current resolution with (bloom_approx_scale_x/viewport_size_x), but + // we don't know the viewport size in this pass. Instead, we'll blur as + // much as it would take to blur away min_allowed_viewport_triads.x. This + // will blur "more than necessary" if the user actually uses more triads, + // but that's not terrible either, because blurring a constant fraction of + // the viewport may better resemble a true optical bloom anyway (since the + // viewport will generally be about the same fraction of each player's + // field of view, regardless of screen size and resolution). + // Assume an extremely large viewport size for asymptotic results. + return bloom_approx_scale_x/max_viewport_size_x * + get_min_sigma_to_blur_triad( + max_viewport_size_x/min_allowed_viewport_triads.x, thresh); +} + +inline float get_center_weight(const float sigma) +{ + // Given a Gaussian blur sigma, get the blur weight for the center texel. + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return get_fast_gaussian_weight_sum_inv(sigma); + #else + const float denom_inv = 0.5/(sigma*sigma); + const float w0 = 1.0; + const float w1 = exp(-1.0 * denom_inv); + const float w2 = exp(-4.0 * denom_inv); + const float w3 = exp(-9.0 * denom_inv); + const float w4 = exp(-16.0 * denom_inv); + const float w5 = exp(-25.0 * denom_inv); + const float w6 = exp(-36.0 * denom_inv); + const float w7 = exp(-49.0 * denom_inv); + const float w8 = exp(-64.0 * denom_inv); + const float w9 = exp(-81.0 * denom_inv); + const float w10 = exp(-100.0 * denom_inv); + const float w11 = exp(-121.0 * denom_inv); + const float w12 = exp(-144.0 * denom_inv); + const float w13 = exp(-169.0 * denom_inv); + const float w14 = exp(-196.0 * denom_inv); + const float w15 = exp(-225.0 * denom_inv); + const float w16 = exp(-256.0 * denom_inv); + const float w17 = exp(-289.0 * denom_inv); + const float w18 = exp(-324.0 * denom_inv); + const float w19 = exp(-361.0 * denom_inv); + const float w20 = exp(-400.0 * denom_inv); + const float w21 = exp(-441.0 * denom_inv); + // Note: If the implementation uses a smaller blur than the max allowed, + // the worst case scenario is that the center weight will be overestimated, + // so we'll put a bit more energy into the brightpass...no huge deal. + // Then again, if the implementation uses a larger blur than the max + // "allowed" because of dynamic branching, the center weight could be + // underestimated, which is more of a problem...consider always using + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // 43x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + + w11 + w12 + w13 + w14 + w15 + w16 + w17 + w18 + w19 + w20 + w21)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + // 31x blur: + const float weight_sum_inv = 1.0 / + (w0 + 2.0 * (w1 + w2 + w3 + w4 + w5 + w6 + w7 + + w8 + w9 + w10 + w11 + w12 + w13 + w14 + w15)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + // 25x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8 + w9 + w10 + w11 + w12)); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + // 17x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * ( + w1 + w2 + w3 + w4 + w5 + w6 + w7 + w8)); + #else + // 9x blur: + const float weight_sum_inv = 1.0 / (w0 + 2.0 * (w1 + w2 + w3 + w4)); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + const float center_weight = weight_sum_inv * weight_sum_inv; + return center_weight; + #endif +} + +inline float3 tex2DblurNfast(const sampler2D texture, const float2 tex_uv, + const float2 dxdy, const float sigma) +{ + // If sigma is static, we can safely branch and use the smallest blur + // that's big enough. Ignore #define hints, because we'll only use a + // large blur if we actually need it, and the branches cost nothing. + #ifndef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #else + // It's still worth branching if the profile supports dynamic branches: + // It's much faster than using a hugely excessive blur, but each branch + // eats ~1% FPS. + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + #endif + #endif + // Failed optimization notes: + // I originally created a same-size mipmapped 5-tap separable blur10 that + // could handle any sigma by reaching into lower mip levels. It was + // as fast as blur25fast for runtime sigmas and a tad faster than + // blur31fast for static sigmas, but mipmapping two viewport-size passes + // ate 10% of FPS across all codepaths, so it wasn't worth it. + #ifdef PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE + if(sigma <= blur9_std_dev) + { + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur17_std_dev) + { + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur25_std_dev) + { + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + } + else if(sigma <= blur31_std_dev) + { + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + } + else + { + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + } + #else + // If we can't afford to branch, we can only guess at what blur + // size we need. Therefore, use the largest blur allowed. + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + return tex2Dblur43fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + return tex2Dblur31fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + return tex2Dblur25fast(texture, tex_uv, dxdy, sigma); + #else + #ifdef PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + return tex2Dblur17fast(texture, tex_uv, dxdy, sigma); + #else + return tex2Dblur9fast(texture, tex_uv, dxdy, sigma); + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + #endif // PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + #endif // PHOSPHOR_BLOOM_BRANCH_FOR_BLUR_SIZE +} + +inline float get_bloom_approx_sigma(const float output_size_x_runtime, + const float estimated_viewport_size_x) +{ + // Requires: 1.) output_size_x_runtime == BLOOM_APPROX.output_size.x. + // This is included for dynamic codepaths just in case the + // following two globals are incorrect: + // 2.) bloom_approx_size_x_for_skip should == the same + // if PHOSPHOR_BLOOM_FAKE is #defined + // 3.) bloom_approx_size_x should == the same otherwise + // Returns: For gaussian4x4, return a dynamic small bloom sigma that's + // as close to optimal as possible given available information. + // For blur3x3, return the a static small bloom sigma that + // works well for typical cases. Otherwise, we're using simple + // bilinear filtering, so use static calculations. + // Assume the default static value. This is a compromise that ensures + // typical triads are blurred, even if unusually large ones aren't. + static const float mask_num_triads_static = + max(min_allowed_viewport_triads.x, mask_num_triads_desired_static); + const float mask_num_triads_from_size = + estimated_viewport_size_x/mask_triad_size_desired; + const float mask_num_triads_runtime = max(min_allowed_viewport_triads.x, + lerp(mask_num_triads_from_size, mask_num_triads_desired, + mask_specify_num_triads)); + // Assume an extremely large viewport size for asymptotic results: + static const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + if(bloom_approx_filter > 1.5) // 4x4 true Gaussian resize + { + // Use the runtime num triads and output size: + const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_runtime; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_runtime/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // account for the Gaussian scanline sigma from the last pass too. + // The bloom will be too wide horizontally but tall enough vertically. + return length(float2(bloom_approx_sigma, beam_max_sigma)); + } + else // 3x3 blur resize (the bilinear resize doesn't need a sigma) + { + // We're either using blur3x3 or bilinear filtering. The biggest + // reason to choose blur3x3 is to avoid dynamic weights, so use a + // static calculation. + #ifdef PHOSPHOR_BLOOM_FAKE + static const float output_size_x_static = + bloom_approx_size_x_for_fake; + #else + static const float output_size_x_static = bloom_approx_size_x; + #endif + static const float asymptotic_triad_size = + max_viewport_size_x/mask_num_triads_static; + const float asymptotic_sigma = get_min_sigma_to_blur_triad( + asymptotic_triad_size, bloom_diff_thresh); + const float bloom_approx_sigma = + asymptotic_sigma * output_size_x_static/max_viewport_size_x; + // The BLOOM_APPROX input has to be ORIG_LINEARIZED to avoid moire, but + // try accounting for the Gaussian scanline sigma from the last pass + // too; use the static default value: + return length(float2(bloom_approx_sigma, beam_max_sigma_static)); + } +} + +inline float get_final_bloom_sigma(const float bloom_sigma_runtime) +{ + // Requires: 1.) bloom_sigma_runtime is a precalculated sigma that's + // optimal for the [known] triad size. + // 2.) Call this from a fragment shader (not a vertex shader), + // or blurring with static sigmas won't be constant-folded. + // Returns: Return the optimistic static sigma if the triad size is + // known at compile time. Otherwise return the optimal runtime + // sigma (10% slower) or an implementation-specific compromise + // between an optimistic or pessimistic static sigma. + // Notes: Call this from the fragment shader, NOT the vertex shader, + // so static sigmas can be constant-folded! + const float bloom_sigma_optimistic = get_min_sigma_to_blur_triad( + mask_triad_size_desired_static, bloom_diff_thresh); + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + return bloom_sigma_runtime; + #else + // Overblurring looks as bad as underblurring, so assume average-size + // triads, not worst-case huge triads: + return bloom_sigma_optimistic; + #endif +} + + +#endif // BLOOM_FUNCTIONS_H + +//////////////////////////// END BLOOM-FUNCTIONS /////////////////////////// + +void main() { + // This pass: Sample (misconverged?) scanlines to the final horizontal + // resolution, apply halation (bouncing electrons), and apply the phosphor + // mask. Fake a bloom if requested. Unless we fake a bloom, the output + // will be dim from the scanline auto-dim, mask dimming, and low gamma. + + // Horizontally sample the current row (a vertically interpolated scanline) + // and account for horizontal convergence offsets, given in units of texels. + const float3 scanline_color_dim = sample_rgb_scanline_horizontal( + VERTICAL_SCANLINEStexture, scanline_tex_uv, + VERTICAL_SCANLINEStexture_size, scanline_texture_size_inv); + const float auto_dim_factor = levels_autodim_temp; + + // Sample the phosphor mask: + const float2 tile_uv_wrap = video_uv * mask_tiles_per_screen; + const float2 mask_tex_uv = convert_phosphor_tile_uv_wrap_to_tex_uv( + tile_uv_wrap, mask_tile_start_uv_and_size); + float3 phosphor_mask_sample; + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + const bool sample_orig_luts = get_mask_sample_mode() > 0.5; + #else + static const bool sample_orig_luts = true; + #endif + if(sample_orig_luts) + { + // If mask_type is static, this branch will be resolved statically. + if(mask_type < 0.5) + { + phosphor_mask_sample = tex2D_linearize( + mask_grille_texture_large, mask_tex_uv).rgb; + } + else if(mask_type < 1.5) + { + phosphor_mask_sample = tex2D_linearize( + mask_slot_texture_large, mask_tex_uv).rgb; + } + else + { + phosphor_mask_sample = tex2D_linearize( + mask_shadow_texture_large, mask_tex_uv).rgb; + } + } + else + { + // Sample the resized mask, and avoid tiling artifacts: + phosphor_mask_sample = tex2Dtiled_mask_linearize( + MASK_RESIZEtexture, mask_tex_uv).rgb; + } + + // Sample the halation texture (auto-dim to match the scanlines), and + // account for both horizontal and vertical convergence offsets, given + // in units of texels horizontally and same-field scanlines vertically: + const float3 halation_color = tex2D_linearize( + HALATION_BLURtexture, halation_tex_uv).rgb; + + // Apply halation: Halation models electrons flying around under the glass + // and hitting the wrong phosphors (of any color). It desaturates, so + // average the halation electrons to a scalar. Reduce the local scanline + // intensity accordingly to conserve energy. + const float3 halation_intensity_dim = + float3(dot(halation_color, float3(auto_dim_factor/3.0))); + const float3 electron_intensity_dim = lerp(scanline_color_dim, + halation_intensity_dim, halation_weight); + + // Apply the phosphor mask: + const float3 phosphor_emission_dim = electron_intensity_dim * + phosphor_mask_sample; + + #ifdef PHOSPHOR_BLOOM_FAKE + // The BLOOM_APPROX pass approximates a blurred version of a masked + // and scanlined image. It's usually used to compute the brightpass, + // but we can also use it to fake the bloom stage entirely. Caveats: + // 1.) A fake bloom is conceptually different, since we're mixing in a + // fully blurred low-res image, and the biggest implication are: + // 2.) If mask_amplify is incorrect, results deteriorate more quickly. + // 3.) The inaccurate blurring hurts quality in high-contrast areas. + // 4.) The bloom_underestimate_levels parameter seems less sensitive. + // Reverse the auto-dimming and amplify to compensate for mask dimming: + #define PHOSPHOR_BLOOM_FAKE_WITH_SIMPLE_BLEND + #ifdef PHOSPHOR_BLOOM_FAKE_WITH_SIMPLE_BLEND + static const float blur_contrast = 1.05; + #else + static const float blur_contrast = 1.0; + #endif + const float mask_amplify = get_mask_amplify(); + const float undim_factor = 1.0/auto_dim_factor; + const float3 phosphor_emission = + phosphor_emission_dim * undim_factor * mask_amplify; + // Get a phosphor blur estimate, accounting for convergence offsets: + const float3 electron_intensity = electron_intensity_dim * undim_factor; + const float3 phosphor_blur_approx_soft = tex2D_linearize( + BLOOM_APPROXtexture, blur3x3_tex_uv).rgb; + const float3 phosphor_blur_approx = lerp(phosphor_blur_approx_soft, + electron_intensity, 0.1) * blur_contrast; + // We could blend between phosphor_emission and phosphor_blur_approx, + // solving for the minimum blend_ratio that avoids clipping past 1.0: + // 1.0 >= total_intensity + // 1.0 >= phosphor_emission * (1.0 - blend_ratio) + + // phosphor_blur_approx * blend_ratio + // blend_ratio = (phosphor_emission - 1.0)/ + // (phosphor_emission - phosphor_blur_approx); + // However, this blurs far more than necessary, because it aims for + // full brightness, not minimal blurring. To fix it, base blend_ratio + // on a max area intensity only so it varies more smoothly: + const float3 phosphor_blur_underestimate = + phosphor_blur_approx * bloom_underestimate_levels; + const float3 area_max_underestimate = + phosphor_blur_underestimate * mask_amplify; + #ifdef PHOSPHOR_BLOOM_FAKE_WITH_SIMPLE_BLEND + const float3 blend_ratio_temp = + (area_max_underestimate - float3(1.0, 1.0, 1.0)) / + (area_max_underestimate - phosphor_blur_underestimate); + #else + // Try doing it like an area-based brightpass. This is nearly + // identical, but it's worth toying with the code in case I ever + // find a way to make it look more like a real bloom. (I've had + // some promising textures from combining an area-based blend ratio + // for the phosphor blur and a more brightpass-like blend-ratio for + // the phosphor emission, but I haven't found a way to make the + // brightness correct across the whole color range, especially with + // different bloom_underestimate_levels values.) + const float desired_triad_size = lerp(mask_triad_size_desired, + output_size.x/mask_num_triads_desired, + mask_specify_num_triads); + const float bloom_sigma = get_min_sigma_to_blur_triad( + desired_triad_size, bloom_diff_thresh); + const float center_weight = get_center_weight(bloom_sigma); + const float3 max_area_contribution_approx = + max(float3(0.0, 0.0, 0.0), phosphor_blur_approx - + center_weight * phosphor_emission); + const float3 area_contrib_underestimate = + bloom_underestimate_levels * max_area_contribution_approx; + const float3 blend_ratio_temp = + ((float3(1.0, 1.0, 1.0) - area_contrib_underestimate) / + area_max_underestimate - float3(1.0, 1.0, 1.0)) / (center_weight - 1.0); + #endif + // Clamp blend_ratio in case it's out-of-range, but be SUPER careful: + // min/max/clamp are BIZARRELY broken with lerp (optimization bug?), + // and this redundant sequence avoids bugs, at least on nVidia cards: + const float3 blend_ratio_clamped = max(clamp(blend_ratio_temp, 0.0, 1.0), 0.0); + const float3 blend_ratio = lerp(blend_ratio_clamped, float3(1.0,1.0,1.0), bloom_excess); + // Blend the blurred and unblurred images: + const float3 phosphor_emission_unclipped = + lerp(phosphor_emission, phosphor_blur_approx, blend_ratio); + // Simulate refractive diffusion by reusing the halation sample. + const float3 pixel_color = lerp(phosphor_emission_unclipped, + halation_color, diffusion_weight); + #else + const float3 pixel_color = phosphor_emission_dim; + #endif + // Encode if necessary, and output. + FragColor = encode_output(float4(pixel_color, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/scanlines-horizontal-apply-mask.vs b/shaders/CRT-Royale.shader/scanlines-horizontal-apply-mask.vs new file mode 100644 index 000000000..41e6f7c18 --- /dev/null +++ b/shaders/CRT-Royale.shader/scanlines-horizontal-apply-mask.vs @@ -0,0 +1,6047 @@ +#version 150 + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 video_uv; + vec2 scanline_tex_uv; + vec2 blur3x3_tex_uv; + vec2 halation_tex_uv; + vec2 scanline_texture_size_inv; + vec4 mask_tile_start_uv_and_size; + vec2 mask_tiles_per_screen; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; +uniform int phase; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +#define VERTICAL_SCANLINEStexture source[5] +#define VERTICAL_SCANLINEStexture_size sourceSize[5].xy +#define VERTICAL_SCANLINESvideo_size sourceSize[5].xy +#define BLOOM_APPROXtexture source[3] +#define BLOOM_APPROXtexture_size sourceSize[3].xy +#define BLOOM_APPROXvideo_size sourceSize[3].xy +#define HALATION_BLURtexture source[1] +#define HALATION_BLURtexture_size sourceSize[1].xy +#define HALATION_BLURvideo_size sourceSize[1].xy +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #define MASK_RESIZEtexture source[0] +#else + #define MASK_RESIZEtexture source[0] +#endif +#define MASK_RESIZEtexture_size sourceSize[0].xy +#define MASK_RESIZEvideo_size sourceSize[0].xy + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +// VERTEX INCLUDES // + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +/////////////////////////////// VERTEX INCLUDES /////////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +//#include "phosphor-mask-resizing.h" + +//////////////////////// BEGIN PHOSPHOR-MASK-RESIZING //////////////////////// + +#ifndef PHOSPHOR_MASK_RESIZING_H +#define PHOSPHOR_MASK_RESIZING_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" +//#include "derived-settings-and-constants.h" + +///////////////////////////// CODEPATH SELECTION ///////////////////////////// + +// Choose a looping strategy based on what's allowed: +// Dynamic loops not allowed: Use a flat static loop. +// Dynamic loops accomodated: Coarsely branch around static loops. +// Dynamic loops assumed allowed: Use a flat dynamic loop. +#ifndef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #ifdef ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + #define BREAK_LOOPS_INTO_PIECES + #else + #define USE_SINGLE_STATIC_LOOP + #endif +#endif // No else needed: Dynamic loops assumed. + + +////////////////////////////////// CONSTANTS ///////////////////////////////// + +// The larger the resized tile, the fewer samples we'll need for downsizing. +// See if we can get a static min tile size > mask_min_allowed_tile_size: +static const float mask_min_allowed_tile_size = ceil( + mask_min_allowed_triad_size * mask_triads_per_tile); +static const float mask_min_expected_tile_size = + mask_min_allowed_tile_size; +// Limit the number of sinc resize taps by the maximum minification factor: +static const float pi_over_lobes = pi/mask_sinc_lobes; +static const float max_sinc_resize_samples_float = 2.0 * mask_sinc_lobes * + mask_resize_src_lut_size.x/mask_min_expected_tile_size; +// Vectorized loops sample in multiples of 4. Round up to be safe: +static const float max_sinc_resize_samples_m4 = ceil( + max_sinc_resize_samples_float * 0.25) * 4.0; + + +///////////////////////// RESAMPLING FUNCTION HELPERS //////////////////////// + +inline float get_dynamic_loop_size(const float magnification_scale) +{ + // Requires: The following global constants must be defined: + // 1.) mask_sinc_lobes + // 2.) max_sinc_resize_samples_m4 + // Returns: The minimum number of texture samples for a correct downsize + // at magnification_scale. + // We're downsizing, so the filter is sized across 2*lobes output pixels + // (not 2*lobes input texels). This impacts distance measurements and the + // minimum number of input samples needed. + const float min_samples_float = 2.0 * mask_sinc_lobes / magnification_scale; + const float min_samples_m4 = ceil(min_samples_float * 0.25) * 4.0; + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + const float max_samples_m4 = max_sinc_resize_samples_m4; + #else // ifdef BREAK_LOOPS_INTO_PIECES + // Simulating loops with branches imposes a 128-sample limit. + const float max_samples_m4 = min(128.0, max_sinc_resize_samples_m4); + #endif + return min(min_samples_m4, max_samples_m4); +} + +float2 get_first_texel_tile_uv_and_dist(const float2 tex_uv, + const float2 tex_size, const float dr, + const float input_tiles_per_texture_r, const float samples, + static const bool vertical) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) input_tiles_per_texture_r is the number of input tiles + // that can fit in the input texture in the direction we're + // resampling this pass. + // 3.) vertical indicates whether we're resampling vertically + // this pass (or horizontally). + // Returns: Pack and return the first sample's tile_uv coord in [0, 1] + // and its texel distance from the destination pixel, in the + // resized dimension only. + // We'll start with the topmost or leftmost sample and work down or right, + // so get the first sample location and distance. Modify both dimensions + // as if we're doing a one-pass 2D resize; we'll throw away the unneeded + // (and incorrect) dimension at the end. + const float2 curr_texel = tex_uv * tex_size; + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 first_texel = prev_texel - float2(samples/2.0 - 1.0); + const float2 first_texel_uv_wrap_2D = first_texel * dr; + const float2 first_texel_dist_2D = curr_texel - first_texel; + // Convert from tex_uv to tile_uv coords so we can sub fracs for fmods. + const float2 first_texel_tile_uv_wrap_2D = + first_texel_uv_wrap_2D * input_tiles_per_texture_r; + // Project wrapped coordinates to the [0, 1] range. We'll do this with all + // samples,but the first texel is special, since it might be negative. + const float2 coord_negative = + float2((first_texel_tile_uv_wrap_2D.x < 0.),(first_texel_tile_uv_wrap_2D.y < 0.)); + const float2 first_texel_tile_uv_2D = + frac(first_texel_tile_uv_wrap_2D) + coord_negative; + // Pack the first texel's tile_uv coord and texel distance in 1D: + const float2 tile_u_and_dist = + float2(first_texel_tile_uv_2D.x, first_texel_dist_2D.x); + const float2 tile_v_and_dist = + float2(first_texel_tile_uv_2D.y, first_texel_dist_2D.y); + return vertical ? tile_v_and_dist : tile_u_and_dist; + //return lerp(tile_u_and_dist, tile_v_and_dist, float(vertical)); +} + +inline float4 tex2Dlod0try(const sampler2D tex, const float2 tex_uv) +{ + // Mipmapping and anisotropic filtering get confused by sinc-resampling. + // One [slow] workaround is to select the lowest mip level: + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + return textureLod(tex, float4(tex_uv, 0.0, 0.0).xy); + #else + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + return tex2Dbias(tex, float4(tex_uv, 0.0, -16.0)); + #else + return texture(tex, tex_uv); + #endif + #endif +} + + +////////////////////////////// LOOP BODY MACROS ////////////////////////////// + +// Using inline functions can exceed the temporary register limit, so we're +// stuck with #define macros (I'm TRULY sorry). They're declared here instead +// of above to be closer to the actual invocation sites. Steps: +// 1.) Get the exact texel location. +// 2.) Sample the phosphor mask (already assumed encoded in linear RGB). +// 3.) Get the distance from the current pixel and sinc weight: +// sinc(dist) = sin(pi * dist)/(pi * dist) +// We can also use the slower/smoother Lanczos instead: +// L(x) = sinc(dist) * sinc(dist / lobes) +// 4.) Accumulate the weight sum in weights, and accumulate the weighted texels +// in pixel_color (we'll normalize outside the loop at the end). +// We vectorize the loop to help reduce the Lanczos window's cost. + + // The r coord is the coord in the dimension we're resizing along (u or v), + // and first_texel_tile_uv_rrrr is a float4 of the first texel's u or v + // tile_uv coord in [0, 1]. tex_uv_r will contain the tile_uv u or v coord + // for four new texel samples. + #define CALCULATE_R_COORD_FOR_4_SAMPLES \ + const float4 true_i = float4(i_base + i) + float4(0.0, 1.0, 2.0, 3.0); \ + const float4 tile_uv_r = frac( \ + first_texel_tile_uv_rrrr + true_i * tile_dr); \ + const float4 tex_uv_r = tile_uv_r * tile_size_uv_r; + + #ifdef PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 pi_dist_over_lobes = pi_over_lobes * dist; \ + const float4 weights = min(sin(pi_dist) * sin(pi_dist_over_lobes) /\ + (pi_dist*pi_dist_over_lobes), float4(1.0)); + #else + #define CALCULATE_SINC_RESAMPLE_WEIGHTS \ + const float4 weights = min(sin(pi_dist)/pi_dist, float4(1.0)); + #endif + + #define UPDATE_COLOR_AND_WEIGHT_SUMS \ + const float4 dist = magnification_scale * \ + abs(first_dist_unscaled - true_i); \ + const float4 pi_dist = pi * dist; \ + CALCULATE_SINC_RESAMPLE_WEIGHTS; \ + pixel_color += new_sample0 * weights.xxx; \ + pixel_color += new_sample1 * weights.yyy; \ + pixel_color += new_sample2 * weights.zzz; \ + pixel_color += new_sample3 * weights.www; \ + weight_sum += weights; + + #define VERTICAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.x)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.z)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv.x, tex_uv_r.w)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + #define HORIZONTAL_SINC_RESAMPLE_LOOP_BODY \ + CALCULATE_R_COORD_FOR_4_SAMPLES; \ + const float3 new_sample0 = tex2Dlod0try(tex, \ + float2(tex_uv_r.x, tex_uv.y)).rgb; \ + const float3 new_sample1 = tex2Dlod0try(tex, \ + float2(tex_uv_r.y, tex_uv.y)).rgb; \ + const float3 new_sample2 = tex2Dlod0try(tex, \ + float2(tex_uv_r.z, tex_uv.y)).rgb; \ + const float3 new_sample3 = tex2Dlod0try(tex, \ + float2(tex_uv_r.w, tex_uv.y)).rgb; \ + UPDATE_COLOR_AND_WEIGHT_SUMS; + + +//////////////////////////// RESAMPLING FUNCTIONS //////////////////////////// + +float3 downsample_vertical_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, static const float dr, + const float magnification_scale, static const float tile_size_uv_r) +{ + // Requires: 1.) dr == du == 1.0/texture_size.x or + // dr == dv == 1.0/texture_size.y + // (whichever direction we're resampling in). + // It's a scalar to save register space. + // 2.) tile_size_uv_r is the number of texels an input tile + // takes up in the input texture, in the direction we're + // resampling this pass. + // 3.) magnification_scale must be <= 1.0. + // Returns: Return a [Lanczos] sinc-resampled pixel of a vertically + // downsized input tile embedded in an input texture. (The + // vertical version is special-cased though: It assumes the + // tile size equals the [static] texture size, since it's used + // on an LUT texture input containing one tile. For more + // generic use, eliminate the "static" in the parameters.) + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dy" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // A static loop can be faster, but it might blur too much from using + // more samples than it should. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along the resized + // dimension) and distance from the output location (in texels): + static const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // true = vertical resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, true); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + static const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + VERTICAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + +float3 downsample_horizontal_sinc_tiled(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, const float dr, + const float magnification_scale, const float tile_size_uv_r) +{ + // Differences from downsample_horizontal_sinc_tiled: + // 1.) The dr and tile_size_uv_r parameters are not static consts. + // 2.) The "vertical" parameter to get_first_texel_tile_uv_and_dist is + // set to false instead of true. + // 3.) The horizontal version of the loop body is used. + // TODO: If we can get guaranteed compile-time dead code elimination, + // we can combine the vertical/horizontal downsampling functions by: + // 1.) Add an extra static const bool parameter called "vertical." + // 2.) Supply it with the result of get_first_texel_tile_uv_and_dist(). + // 3.) Use a conditional assignment in the loop body macro. This is the + // tricky part: We DO NOT want to incur the extra conditional + // assignment in the inner loop at runtime! + // The "r" in "dr," "tile_size_uv_r," etc. refers to the dimension + // we're resizing along, e.g. "dx" in this case. + #ifdef USE_SINGLE_STATIC_LOOP + // If we have to load all samples, we might as well use them. + static const int samples = int(max_sinc_resize_samples_m4); + #else + const int samples = int(get_dynamic_loop_size(magnification_scale)); + #endif + + // Get the first sample location (scalar tile uv coord along resized + // dimension) and distance from the output location (in texels): + const float input_tiles_per_texture_r = 1.0/tile_size_uv_r; + // false = horizontal resize: + const float2 first_texel_tile_r_and_dist = get_first_texel_tile_uv_and_dist( + tex_uv, tex_size, dr, input_tiles_per_texture_r, samples, false); + const float4 first_texel_tile_uv_rrrr = first_texel_tile_r_and_dist.xxxx; + const float4 first_dist_unscaled = first_texel_tile_r_and_dist.yyyy; + // Get the tile sample offset: + const float tile_dr = dr * input_tiles_per_texture_r; + + // Sum up each weight and weighted sample color, varying the looping + // strategy based on our expected dynamic loop capabilities. See the + // loop body macros above. + int i_base = 0; + float4 weight_sum = float4(0.0); + float3 pixel_color = float3(0.0); + static const int i_step = 4; + #ifdef BREAK_LOOPS_INTO_PIECES + if(samples - i_base >= 64) + { + for(int i = 0; i < 64; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 64; + } + if(samples - i_base >= 32) + { + for(int i = 0; i < 32; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 32; + } + if(samples - i_base >= 16) + { + for(int i = 0; i < 16; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 16; + } + if(samples - i_base >= 8) + { + for(int i = 0; i < 8; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 8; + } + if(samples - i_base >= 4) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + i_base += 4; + } + // Do another 4-sample block for a total of 128 max samples. + if(samples - i_base > 0) + { + for(int i = 0; i < 4; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + } + #else + for(int i = 0; i < samples; i += i_step) + { + HORIZONTAL_SINC_RESAMPLE_LOOP_BODY; + } + #endif + // Normalize so the weight_sum == 1.0, and return: + const float2 weight_sum_reduce = weight_sum.xy + weight_sum.zw; + const float3 scalar_weight_sum = float3(weight_sum_reduce.x + + weight_sum_reduce.y); + return (pixel_color/scalar_weight_sum); +} + + +//////////////////////////// TILE SIZE CALCULATION /////////////////////////// + +float2 get_resized_mask_tile_size(const float2 estimated_viewport_size, + const float2 estimated_mask_resize_output_size, + const bool solemnly_swear_same_inputs_for_every_pass) +{ + // Requires: The following global constants must be defined according to + // certain constraints: + // 1.) mask_resize_num_triads: Must be high enough that our + // mask sampling method won't have artifacts later + // (long story; see derived-settings-and-constants.h) + // 2.) mask_resize_src_lut_size: Texel size of our mask LUT + // 3.) mask_triads_per_tile: Num horizontal triads in our LUT + // 4.) mask_min_allowed_triad_size: User setting (the more + // restrictive it is, the faster the resize will go) + // 5.) mask_min_allowed_tile_size_x < mask_resize_src_lut_size.x + // 6.) mask_triad_size_desired_{runtime, static} + // 7.) mask_num_triads_desired_{runtime, static} + // 8.) mask_specify_num_triads must be 0.0/1.0 (false/true) + // The function parameters must be defined as follows: + // 1.) estimated_viewport_size == (final viewport size); + // If mask_specify_num_triads is 1.0/true and the viewport + // estimate is wrong, the number of triads will differ from + // the user's preference by about the same factor. + // 2.) estimated_mask_resize_output_size: Must equal the + // output size of the MASK_RESIZE pass. + // Exception: The x component may be estimated garbage if + // and only if the caller throws away the x result. + // 3.) solemnly_swear_same_inputs_for_every_pass: Set to false, + // unless you can guarantee that every call across every + // pass will use the same sizes for the other parameters. + // When calling this across multiple passes, always use the + // same y viewport size/scale, and always use the same x + // viewport size/scale when using the x result. + // Returns: Return the final size of a manually resized mask tile, after + // constraining the desired size to avoid artifacts. Under + // unusual circumstances, tiles may become stretched vertically + // (see wall of text below). + // Stated tile properties must be correct: + static const float tile_aspect_ratio_inv = + mask_resize_src_lut_size.y/mask_resize_src_lut_size.x; + static const float tile_aspect_ratio = 1.0/tile_aspect_ratio_inv; + static const float2 tile_aspect = float2(1.0, tile_aspect_ratio_inv); + // If mask_specify_num_triads is 1.0/true and estimated_viewport_size.x is + // wrong, the user preference will be misinterpreted: + const float desired_tile_size_x = mask_triads_per_tile * lerp( + mask_triad_size_desired, + estimated_viewport_size.x / mask_num_triads_desired, + mask_specify_num_triads); + if(get_mask_sample_mode() > 0.5) + { + // We don't need constraints unless we're sampling MASK_RESIZE. + return desired_tile_size_x * tile_aspect; + } + // Make sure we're not upsizing: + const float temp_tile_size_x = + min(desired_tile_size_x, mask_resize_src_lut_size.x); + // Enforce min_tile_size and max_tile_size in both dimensions: + const float2 temp_tile_size = temp_tile_size_x * tile_aspect; + static const float2 min_tile_size = + mask_min_allowed_tile_size * tile_aspect; + const float2 max_tile_size = + estimated_mask_resize_output_size / mask_resize_num_tiles; + const float2 clamped_tile_size = + clamp(temp_tile_size, min_tile_size, max_tile_size); + // Try to maintain tile_aspect_ratio. This is the tricky part: + // If we're currently resizing in the y dimension, the x components + // could be MEANINGLESS. (If estimated_mask_resize_output_size.x is + // bogus, then so is max_tile_size.x and clamped_tile_size.x.) + // We can't adjust the y size based on clamped_tile_size.x. If it + // clamps when it shouldn't, it won't clamp again when later passes + // call this function with the correct sizes, and the discrepancy will + // break the sampling coords in MASKED_SCANLINES. Instead, we'll limit + // the x size based on the y size, but not vice versa, unless the + // caller swears the parameters were the same (correct) in every pass. + // As a result, triads could appear vertically stretched if: + // a.) mask_resize_src_lut_size.x > mask_resize_src_lut_size.y: Wide + // LUT's might clamp x more than y (all provided LUT's are square) + // b.) true_viewport_size.x < true_viewport_size.y: The user is playing + // with a vertically oriented screen (not accounted for anyway) + // c.) mask_resize_viewport_scale.x < masked_resize_viewport_scale.y: + // Viewport scales are equal by default. + // If any of these are the case, you can fix the stretching by setting: + // mask_resize_viewport_scale.x = mask_resize_viewport_scale.y * + // (1.0 / min_expected_aspect_ratio) * + // (mask_resize_src_lut_size.x / mask_resize_src_lut_size.y) + const float x_tile_size_from_y = + clamped_tile_size.y * tile_aspect_ratio; + const float y_tile_size_from_x = lerp(clamped_tile_size.y, + clamped_tile_size.x * tile_aspect_ratio_inv, + float(solemnly_swear_same_inputs_for_every_pass)); + const float2 reclamped_tile_size = float2( + min(clamped_tile_size.x, x_tile_size_from_y), + min(clamped_tile_size.y, y_tile_size_from_x)); + // We need integer tile sizes in both directions for tiled sampling to + // work correctly. Use floor (to make sure we don't round up), but be + // careful to avoid a rounding bug where floor decreases whole numbers: + const float2 final_resized_tile_size = + floor(reclamped_tile_size + float2(FIX_ZERO(0.0))); + return final_resized_tile_size; +} + + +///////////////////////// FINAL MASK SAMPLING HELPERS //////////////////////// + +float4 get_mask_sampling_parameters(const float2 mask_resize_texture_size, + const float2 mask_resize_video_size, const float2 true_viewport_size, + out float2 mask_tiles_per_screen) +{ + // Requires: 1.) Requirements of get_resized_mask_tile_size() must be + // met, particularly regarding global constants. + // The function parameters must be defined as follows: + // 1.) mask_resize_texture_size == MASK_RESIZE.texture_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 2.) mask_resize_video_size == MASK_RESIZE.video_size + // if get_mask_sample_mode() is 0 (otherwise anything) + // 3.) true_viewport_size == output_size for a pass set to + // 1.0 viewport scale (i.e. it must be correct) + // Returns: Return a float4 containing: + // xy: tex_uv coords for the start of the mask tile + // zw: tex_uv size of the mask tile from start to end + // mask_tiles_per_screen is an out parameter containing the + // number of mask tiles that will fit on the screen. + // First get the final resized tile size. The viewport size and mask + // resize viewport scale must be correct, but don't solemnly swear they + // were correct in both mask resize passes unless you know it's true. + // (We can better ensure a correct tile aspect ratio if the parameters are + // guaranteed correct in all passes...but if we lie, we'll get inconsistent + // sizes across passes, resulting in broken texture coordinates.) + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_tile_size = get_resized_mask_tile_size( + true_viewport_size, mask_resize_video_size, false); + if(mask_sample_mode < 0.5) + { + // Sample MASK_RESIZE: The resized tile is a fraction of the texture + // size and starts at a nonzero offset to allow for border texels: + const float2 mask_tile_uv_size = mask_resize_tile_size / + mask_resize_texture_size; + const float2 skipped_tiles = mask_start_texels/mask_resize_tile_size; + const float2 mask_tile_start_uv = skipped_tiles * mask_tile_uv_size; + // mask_tiles_per_screen must be based on the *true* viewport size: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + return float4(mask_tile_start_uv, mask_tile_uv_size); + } + else + { + // If we're tiling at the original size (1:1 pixel:texel), redefine a + // "tile" to be the full texture containing many triads. Otherwise, + // we're hardware-resampling an LUT, and the texture truly contains a + // single unresized phosphor mask tile anyway. + static const float2 mask_tile_uv_size = float2(1.0); + static const float2 mask_tile_start_uv = float2(0.0); + if(mask_sample_mode > 1.5) + { + // Repeat the full LUT at a 1:1 pixel:texel ratio without resizing: + mask_tiles_per_screen = true_viewport_size/mask_texture_large_size; + } + else + { + // Hardware-resize the original LUT: + mask_tiles_per_screen = true_viewport_size / mask_resize_tile_size; + } + return float4(mask_tile_start_uv, mask_tile_uv_size); + } +} +/* +float2 fix_tiling_discontinuities_normalized(const float2 tile_uv, + float2 duv_dx, float2 duv_dy) +{ + // Requires: 1.) duv_dx == ddx(tile_uv) + // 2.) duv_dy == ddy(tile_uv) + // 3.) tile_uv contains tile-relative uv coords in [0, 1], + // such that (0.5, 0.5) is the center of a tile, etc. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // Returns: Return new tile_uv coords that contain no discontinuities + // across a 2x2 pixel quad. + // Description: + // When uv coords wrap from 1.0 to 0.0, they create a discontinuity in the + // derivatives, which we assume happened if the absolute difference between + // any fragment in a 2x2 block is > ~half a tile. If the current block has + // a u or v discontinuity and the current fragment is in the first half of + // the tile along that axis (i.e. it wrapped from 1.0 to 0.0), add a tile + // to that coord to make the 2x2 block continuous. (It will now have a + // coord > 1.0 in the padding area beyond the tile.) This function takes + // derivatives as parameters so the caller can reuse them. + // In case we're using high-quality (nVidia-style) derivatives, ensure + // diagonically opposite fragments see each other for correctness: + duv_dx = abs(duv_dx) + abs(ddy(duv_dx)); + duv_dy = abs(duv_dy) + abs(ddx(duv_dy)); + const float2 pixel_in_first_half_tile = float2((tile_uv.x < 0.5),(tile_uv.y < 0.5)); + const float2 jump_exists = float2(((duv_dx + duv_dy).x > 0.5),((duv_dx + duv_dy).y > 0.5)); + return tile_uv + jump_exists * pixel_in_first_half_tile; +} +*/ +float2 convert_phosphor_tile_uv_wrap_to_tex_uv(const float2 tile_uv_wrap, + const float4 mask_tile_start_uv_and_size) +{ + // Requires: 1.) tile_uv_wrap contains tile-relative uv coords, where the + // tile spans from [0, 1], such that (0.5, 0.5) is at the + // tile center. The input coords can range from [0, inf], + // and their fractional parts map to a repeated tile. + // ("Tile" can mean texture, the video embedded in the + // texture, or some other "tile" embedded in a texture.) + // 2.) mask_tile_start_uv_and_size.xy contains tex_uv coords + // for the start of the embedded tile in the full texture. + // 3.) mask_tile_start_uv_and_size.zw contains the [fractional] + // tex_uv size of the embedded tile in the full texture. + // Returns: Return tex_uv coords (used for texture sampling) + // corresponding to tile_uv_wrap. + if(get_mask_sample_mode() < 0.5) + { + // Manually repeat the resized mask tile to fill the screen: + // First get fractional tile_uv coords. Using frac/fmod on coords + // confuses anisotropic filtering; fix it as user options dictate. + // derived-settings-and-constants.h disables incompatible options. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + float2 tile_uv = frac(tile_uv_wrap * 0.5) * 2.0; + #else + float2 tile_uv = frac(tile_uv_wrap); + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + const float2 tile_uv_dx = ddx(tile_uv); + const float2 tile_uv_dy = ddy(tile_uv); + tile_uv = fix_tiling_discontinuities_normalized(tile_uv, + tile_uv_dx, tile_uv_dy); + #endif + // The tile is embedded in a padded FBO, and it may start at a + // nonzero offset if border texels are used to avoid artifacts: + const float2 mask_tex_uv = mask_tile_start_uv_and_size.xy + + tile_uv * mask_tile_start_uv_and_size.zw; + return mask_tex_uv; + } + else + { + // Sample from the input phosphor mask texture with hardware tiling. + // If we're tiling at the original size (mode 2), the "tile" is the + // whole texture, and it contains a large number of triads mapped with + // a 1:1 pixel:texel ratio. OTHERWISE, the texture contains a single + // unresized tile. tile_uv_wrap already has correct coords for both! + return tile_uv_wrap; + } +} + + +#endif // PHOSPHOR_MASK_RESIZING_H + +///////////////////////// END PHOSPHOR-MASK-RESIZING ///////////////////////// + +//#include "../../../../include/gamma-management.h" +// already got it + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////////// HELPERS ////////////////////////////////// + +inline float4 tex2Dtiled_mask_linearize(const sampler2D tex, + const float2 tex_uv) +{ + // If we're manually tiling a texture, anisotropic filtering can get + // confused. One workaround is to just select the lowest mip level: + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + // TODO: Use tex2Dlod_linearize with a calculated mip level. + return tex2Dlod_linearize(tex, float4(tex_uv, 0.0, 0.0)); + #else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + return tex2Dbias_linearize(tex, float4(tex_uv, 0.0, -16.0)); + #else + return tex2D_linearize(tex, tex_uv); + #endif + #endif + #else + return tex2D_linearize(tex, tex_uv); + #endif +} + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +// END VERTEX INCLUDES // + +float bloom_approx_scale_x = targetSize.x / sourceSize[0].y; +const float max_viewport_size_x = 1080.0*1024.0*(4.0/3.0); + +void main() { + gl_Position = position; + vTexCoord = texCoord * 1.00001; + + float2 tex_uv = vTexCoord.xy; + // Our various input textures use different coords. + video_uv = tex_uv * texture_size/video_size; + scanline_texture_size_inv = + float2(1.0, 1.0)/VERTICAL_SCANLINEStexture_size; + //video_uv = video_uv; + scanline_tex_uv = video_uv * VERTICAL_SCANLINESvideo_size * + scanline_texture_size_inv; + blur3x3_tex_uv = video_uv * BLOOM_APPROXvideo_size / + BLOOM_APPROXtexture_size; + halation_tex_uv = video_uv * HALATION_BLURvideo_size / + HALATION_BLURtexture_size; + //scanline_texture_size_inv = scanline_texture_size_inv; + + // Get a consistent name for the final mask texture size. Sample mode 0 + // uses the manually resized mask, but ignore it if we never resized. + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + const float mask_sample_mode = get_mask_sample_mode(); + const float2 mask_resize_texture_size = mask_sample_mode < 0.5 ? + MASK_RESIZEtexture_size : mask_texture_large_size; + const float2 mask_resize_video_size = mask_sample_mode < 0.5 ? + MASK_RESIZEvideo_size : mask_texture_large_size; + #else + const float2 mask_resize_texture_size = mask_texture_large_size; + const float2 mask_resize_video_size = mask_texture_large_size; + #endif + // Compute mask tile dimensions, starting points, etc.: + //float2 mask_tiles_per_screen; + mask_tile_start_uv_and_size = get_mask_sampling_parameters( + mask_resize_texture_size, mask_resize_video_size, output_size, + mask_tiles_per_screen); + //mask_tiles_per_screen = mask_tiles_per_screen; +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/scanlines-vertical-interlacing.fs b/shaders/CRT-Royale.shader/scanlines-vertical-interlacing.fs new file mode 100644 index 000000000..d090c5297 --- /dev/null +++ b/shaders/CRT-Royale.shader/scanlines-vertical-interlacing.fs @@ -0,0 +1,5963 @@ +#version 150 + +uniform sampler2D source[]; +uniform vec4 sourceSize[]; +uniform vec4 targetSize; +uniform int phase; + +in Vertex { + vec2 vTexCoord; + vec2 uv_step; + vec2 il_step_multiple; + float pixel_height_in_scanlines; +}; + +out vec4 FragColor; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 0.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-params.h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + vec2 tex_uv = vTexCoord.xy; + // This pass: Sample multiple (misconverged?) scanlines to the final + // vertical resolution. Temporarily auto-dim the output to avoid clipping. + + // Read some attributes into local variables: + float2 texture_size_ = texture_size; + float2 texture_size_inv = 1.0/texture_size_; + //const float2 uv_step = uv_step; + //const float2 il_step_multiple = il_step_multiple; + float frame_count = float(frame_count); + const float ph = pixel_height_in_scanlines; + + // Get the uv coords of the previous scanline (in this field), and the + // scanline's distance from this sample, in scanlines. + float dist; + const float2 scanline_uv = get_last_scanline_uv(tex_uv, texture_size_, + texture_size_inv, il_step_multiple, frame_count, dist); + // Consider 2, 3, 4, or 6 scanlines numbered 0-5: The previous and next + // scanlines are numbered 2 and 3. Get scanline colors colors (ignore + // horizontal sampling, since since output_size.x = video_size.x). + // NOTE: Anisotropic filtering creates interlacing artifacts, which is why + // ORIG_LINEARIZED bobbed any interlaced input before this pass. + const float2 v_step = float2(0.0, uv_step.y); + const float3 scanline2_color = tex2D_linearize(input_texture, scanline_uv).rgb; + const float3 scanline3_color = + tex2D_linearize(input_texture, scanline_uv + v_step).rgb; + float3 scanline0_color, scanline1_color, scanline4_color, scanline5_color, + scanline_outside_color; + float dist_round; + // Use scanlines 0, 1, 4, and 5 for a total of 6 scanlines: + if(beam_num_scanlines > 5.5) + { + scanline1_color = + tex2D_linearize(input_texture, scanline_uv - v_step).rgb; + scanline4_color = + tex2D_linearize(input_texture, scanline_uv + 2.0 * v_step).rgb; + scanline0_color = + tex2D_linearize(input_texture, scanline_uv - 2.0 * v_step).rgb; + scanline5_color = + tex2D_linearize(input_texture, scanline_uv + 3.0 * v_step).rgb; + } + // Use scanlines 1, 4, and either 0 or 5 for a total of 5 scanlines: + else if(beam_num_scanlines > 4.5) + { + scanline1_color = + tex2D_linearize(input_texture, scanline_uv - v_step).rgb; + scanline4_color = + tex2D_linearize(input_texture, scanline_uv + 2.0 * v_step).rgb; + // dist is in [0, 1] + dist_round = round(dist); + const float2 sample_0_or_5_uv_off = + lerp(-2.0 * v_step, 3.0 * v_step, dist_round); + // Call this "scanline_outside_color" to cope with the conditional + // scanline number: + scanline_outside_color = tex2D_linearize( + input_texture, scanline_uv + sample_0_or_5_uv_off).rgb; + } + // Use scanlines 1 and 4 for a total of 4 scanlines: + else if(beam_num_scanlines > 3.5) + { + scanline1_color = + tex2D_linearize(input_texture, scanline_uv - v_step).rgb; + scanline4_color = + tex2D_linearize(input_texture, scanline_uv + 2.0 * v_step).rgb; + } + // Use scanline 1 or 4 for a total of 3 scanlines: + else if(beam_num_scanlines > 2.5) + { + // dist is in [0, 1] + dist_round = round(dist); + const float2 sample_1or4_uv_off = + lerp(-v_step, 2.0 * v_step, dist_round); + scanline_outside_color = tex2D_linearize( + input_texture, scanline_uv + sample_1or4_uv_off).rgb; + } + + // Compute scanline contributions, accounting for vertical convergence. + // Vertical convergence offsets are in units of current-field scanlines. + // dist2 means "positive sample distance from scanline 2, in scanlines:" + float3 dist2 = float3(dist); + if(beam_misconvergence) + { + const float3 convergence_offsets_vert_rgb = + get_convergence_offsets_y_vector(); + dist2 = float3(dist) - convergence_offsets_vert_rgb; + } + // Calculate {sigma, shape}_range outside of scanline_contrib so it's only + // done once per pixel (not 6 times) with runtime params. Don't reuse the + // vertex shader calculations, so static versions can be constant-folded. + const float sigma_range = max(beam_max_sigma, beam_min_sigma) - + beam_min_sigma; + const float shape_range = max(beam_max_shape, beam_min_shape) - + beam_min_shape; + // Calculate and sum final scanline contributions, starting with lines 2/3. + // There is no normalization step, because we're not interpolating a + // continuous signal. Instead, each scanline is an additive light source. + const float3 scanline2_contrib = scanline_contrib(dist2, + scanline2_color, ph, sigma_range, shape_range); + const float3 scanline3_contrib = scanline_contrib(abs(float3(1.0,1.0,1.0) - dist2), + scanline3_color, ph, sigma_range, shape_range); + float3 scanline_intensity = scanline2_contrib + scanline3_contrib; + if(beam_num_scanlines > 5.5) + { + const float3 scanline0_contrib = + scanline_contrib(dist2 + float3(2.0,2.0,2.0), scanline0_color, + ph, sigma_range, shape_range); + const float3 scanline1_contrib = + scanline_contrib(dist2 + float3(1.0,1.0,1.0), scanline1_color, + ph, sigma_range, shape_range); + const float3 scanline4_contrib = + scanline_contrib(abs(float3(2.0,2.0,2.0) - dist2), scanline4_color, + ph, sigma_range, shape_range); + const float3 scanline5_contrib = + scanline_contrib(abs(float3(3.0) - dist2), scanline5_color, + ph, sigma_range, shape_range); + scanline_intensity += scanline0_contrib + scanline1_contrib + + scanline4_contrib + scanline5_contrib; + } + else if(beam_num_scanlines > 4.5) + { + const float3 scanline1_contrib = + scanline_contrib(dist2 + float3(1.0,1.0,1.0), scanline1_color, + ph, sigma_range, shape_range); + const float3 scanline4_contrib = + scanline_contrib(abs(float3(2.0,2.0,2.0) - dist2), scanline4_color, + ph, sigma_range, shape_range); + const float3 dist0or5 = lerp( + dist2 + float3(2.0,2.0,2.0), float3(3.0,3.0,3.0) - dist2, dist_round); + const float3 scanline0or5_contrib = scanline_contrib( + dist0or5, scanline_outside_color, ph, sigma_range, shape_range); + scanline_intensity += scanline1_contrib + scanline4_contrib + + scanline0or5_contrib; + } + else if(beam_num_scanlines > 3.5) + { + const float3 scanline1_contrib = + scanline_contrib(dist2 + float3(1.0,1.0,1.0), scanline1_color, + ph, sigma_range, shape_range); + const float3 scanline4_contrib = + scanline_contrib(abs(float3(2.0,2.0,2.0) - dist2), scanline4_color, + ph, sigma_range, shape_range); + scanline_intensity += scanline1_contrib + scanline4_contrib; + } + else if(beam_num_scanlines > 2.5) + { + const float3 dist1or4 = lerp( + dist2 + float3(1.0,1.0,1.0), float3(2.0,2.0,2.0) - dist2, dist_round); + const float3 scanline1or4_contrib = scanline_contrib( + dist1or4, scanline_outside_color, ph, sigma_range, shape_range); + scanline_intensity += scanline1or4_contrib; + } + + // Auto-dim the image to avoid clipping, encode if necessary, and output. + // My original idea was to compute a minimal auto-dim factor and put it in + // the alpha channel, but it wasn't working, at least not reliably. This + // is faster anyway, levels_autodim_temp = 0.5 isn't causing banding. + FragColor = encode_output(float4(scanline_intensity * levels_autodim_temp, 1.0)); +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/scanlines-vertical-interlacing.vs b/shaders/CRT-Royale.shader/scanlines-vertical-interlacing.vs new file mode 100644 index 000000000..8fe7b14c1 --- /dev/null +++ b/shaders/CRT-Royale.shader/scanlines-vertical-interlacing.vs @@ -0,0 +1,5830 @@ +#version 150 + +in vec4 position; +in vec2 texCoord; + +out Vertex { + vec2 vTexCoord; + vec2 uv_step; + vec2 il_step_multiple; + float pixel_height_in_scanlines; +}; + +uniform vec4 targetSize; +uniform vec4 sourceSize[]; + +// USER SETTINGS BLOCK // + +#define crt_gamma 2.500000 +#define lcd_gamma 2.200000 +#define levels_contrast 1.0 +#define halation_weight 0.0 +#define diffusion_weight 0.075 +#define bloom_underestimate_levels 0.8 +#define bloom_excess 0.000000 +#define beam_min_sigma 0.020000 +#define beam_max_sigma 0.300000 +#define beam_spot_power 0.330000 +#define beam_min_shape 2.000000 +#define beam_max_shape 4.000000 +#define beam_shape_power 0.250000 +#define beam_horiz_filter 0.000000 +#define beam_horiz_sigma 0.35 +#define beam_horiz_linear_rgb_weight 1.000000 +#define convergence_offset_x_r -0.000000 +#define convergence_offset_x_g 0.000000 +#define convergence_offset_x_b 0.000000 +#define convergence_offset_y_r 0.000000 +#define convergence_offset_y_g -0.000000 +#define convergence_offset_y_b 0.000000 +#define mask_type 1.000000 +#define mask_sample_mode_desired 0.000000 +#define mask_specify_num_triads 0.000000 +#define mask_triad_size_desired 3.000000 +#define mask_num_triads_desired 480.000000 +#define aa_subpixel_r_offset_x_runtime -0.0 +#define aa_subpixel_r_offset_y_runtime 0.000000 +#define aa_cubic_c 0.500000 +#define aa_gauss_sigma 0.500000 +#define geom_mode_runtime 2.000000 +#define geom_radius 2.000000 +#define geom_view_dist 2.000000 +#define geom_tilt_angle_x 0.000000 +#define geom_tilt_angle_y 0.000000 +#define geom_aspect_ratio_x 432.000000 +#define geom_aspect_ratio_y 329.000000 +#define geom_overscan_x 1.000000 +#define geom_overscan_y 1.000000 +#define border_size 0.015 +#define border_darkness 2.0 +#define border_compress 2.500000 +#define interlace_bff 0.000000 +#define interlace_1080i 0.000000 + +// END USER SETTINGS BLOCK // + +// compatibility macros for transparently converting HLSLisms into GLSLisms +#define mul(a,b) (b*a) +#define lerp(a,b,c) mix(a,b,c) +#define saturate(c) clamp(c, 0.0, 1.0) +#define frac(x) (fract(x)) +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#define float2x2 mat2x2 +#define float3x3 mat3x3 +#define float4x4 mat4x4 +#define float4x3 mat4x3 +#define float2x4 mat2x4 +#define IN params +#define texture_size sourceSize[0].xy +#define video_size sourceSize[0].xy +#define output_size targetSize.xy +#define frame_count phase +#define static +#define inline +#define const +#define fmod(x,y) mod(x,y) +#define ddx(c) dFdx(c) +#define ddy(c) dFdy(c) +#define atan2(x,y) atan(y,x) +#define rsqrt(c) inversesqrt(c) + +#define input_texture source[0] + +#if defined(GL_ES) + #define COMPAT_PRECISION mediump +#else + #define COMPAT_PRECISION +#endif + +#if __VERSION__ >= 130 + #define COMPAT_TEXTURE texture +#else + #define COMPAT_TEXTURE texture2D +#endif + +////////////////////////////////// INCLUDES ////////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "bind-shader-params.h" + +///////////////////////////// BEGIN BIND-SHADER-PARAMS //////////////////////////// + +#ifndef BIND_SHADER_PARAMS_H +#define BIND_SHADER_PARAMS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "derived-settings-and-constants.h" + +///////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +//////////////////// END DERIVED-SETTINGS-AND-CONSTANTS ///////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +// Override some parameters for gamma-management.h and tex2Dantialias.h: +#define OVERRIDE_DEVICE_GAMMA +static const float gba_gamma = 3.5; // Irrelevant but necessary to define. +#define ANTIALIAS_OVERRIDE_BASICS +#define ANTIALIAS_OVERRIDE_PARAMETERS + +// Provide accessors for vector constants that pack scalar uniforms: +inline float2 get_aspect_vector(const float geom_aspect_ratio) +{ + // Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent + // the absolute scale from affecting the uv-mapping for curvature: + const float geom_clamped_aspect_ratio = + min(geom_aspect_ratio, geom_max_aspect_ratio); + const float2 geom_aspect = + normalize(float2(geom_clamped_aspect_ratio, 1.0)); + return geom_aspect; +} + +inline float2 get_geom_overscan_vector() +{ + return float2(geom_overscan_x, geom_overscan_y); +} + +inline float2 get_geom_tilt_angle_vector() +{ + return float2(geom_tilt_angle_x, geom_tilt_angle_y); +} + +inline float3 get_convergence_offsets_x_vector() +{ + return float3(convergence_offset_x_r, convergence_offset_x_g, + convergence_offset_x_b); +} + +inline float3 get_convergence_offsets_y_vector() +{ + return float3(convergence_offset_y_r, convergence_offset_y_g, + convergence_offset_y_b); +} + +inline float2 get_convergence_offsets_r_vector() +{ + return float2(convergence_offset_x_r, convergence_offset_y_r); +} + +inline float2 get_convergence_offsets_g_vector() +{ + return float2(convergence_offset_x_g, convergence_offset_y_g); +} + +inline float2 get_convergence_offsets_b_vector() +{ + return float2(convergence_offset_x_b, convergence_offset_y_b); +} + +inline float2 get_aa_subpixel_r_offset() +{ + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + // WARNING: THIS IS EXTREMELY EXPENSIVE. + return float2(aa_subpixel_r_offset_x_runtime, + aa_subpixel_r_offset_y_runtime); + #else + return aa_subpixel_r_offset_static; + #endif + #else + return aa_subpixel_r_offset_static; + #endif +} + +// Provide accessors settings which still need "cooking:" +inline float get_mask_amplify() +{ + static const float mask_grille_amplify = 1.0/mask_grille_avg_color; + static const float mask_slot_amplify = 1.0/mask_slot_avg_color; + static const float mask_shadow_amplify = 1.0/mask_shadow_avg_color; + return mask_type < 0.5 ? mask_grille_amplify : + mask_type < 1.5 ? mask_slot_amplify : + mask_shadow_amplify; +} + +inline float get_mask_sample_mode() +{ + #ifdef RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_desired; + #else + return clamp(mask_sample_mode_desired, 1.0, 2.0); + #endif + #else + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + return mask_sample_mode_static; + #else + return clamp(mask_sample_mode_static, 1.0, 2.0); + #endif + #endif +} + +#endif // BIND_SHADER_PARAMS_H + +//////////////////////////// END BIND-SHADER-PARAMS /////////////////////////// + +//#include "scanline-functions.h" + +///////////////////////////// BEGIN SCANLINE-FUNCTIONS //////////////////////////// + +#ifndef SCANLINE_FUNCTIONS_H +#define SCANLINE_FUNCTIONS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +//////////////////////////// END USER-SETTINGS ////////////////////////// + +//#include "derived-settings-and-constants.h" + +//////////////////// BEGIN DERIVED-SETTINGS-AND-CONSTANTS //////////////////// + +#ifndef DERIVED_SETTINGS_AND_CONSTANTS_H +#define DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// GPL LICENSE NOTICE ///////////////////////////// + +// crt-royale: A full-featured CRT shader, with cheese. +// Copyright (C) 2014 TroggleMonkey +// +// This program is free software; you can redistribute it and/or modify it +// under the terms of the GNU General Public License as published by the Free +// Software Foundation; either version 2 of the License, or any later version. +// +// This program is distributed in the hope that it will be useful, but WITHOUT +// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +// more details. +// +// You should have received a copy of the GNU General Public License along with +// this program; if not, write to the Free Software Foundation, Inc., 59 Temple +// Place, Suite 330, Boston, MA 02111-1307 USA + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// These macros and constants can be used across the whole codebase. +// Unlike the values in user-settings.cgh, end users shouldn't modify these. + + +/////////////////////////////// BEGIN INCLUDES /////////////////////////////// + +//#include "../user-settings.h" + +///////////////////////////// BEGIN USER-SETTINGS //////////////////////////// + +#ifndef USER_SETTINGS_H +#define USER_SETTINGS_H + +///////////////////////////// DRIVER CAPABILITIES //////////////////////////// + +// The Cg compiler uses different "profiles" with different capabilities. +// This shader requires a Cg compilation profile >= arbfp1, but a few options +// require higher profiles like fp30 or fp40. The shader can't detect profile +// or driver capabilities, so instead you must comment or uncomment the lines +// below with "//" before "#define." Disable an option if you get compilation +// errors resembling those listed. Generally speaking, all of these options +// will run on nVidia cards, but only DRIVERS_ALLOW_TEX2DBIAS (if that) is +// likely to run on ATI/AMD, due to the Cg compiler's profile limitations. + +// Derivatives: Unsupported on fp20, ps_1_1, ps_1_2, ps_1_3, and arbfp1. +// Among other things, derivatives help us fix anisotropic filtering artifacts +// with curved manually tiled phosphor mask coords. Related errors: +// error C3004: function "float2 ddx(float2);" not supported in this profile +// error C3004: function "float2 ddy(float2);" not supported in this profile + //#define DRIVERS_ALLOW_DERIVATIVES + +// Fine derivatives: Unsupported on older ATI cards. +// Fine derivatives enable 2x2 fragment block communication, letting us perform +// fast single-pass blur operations. If your card uses coarse derivatives and +// these are enabled, blurs could look broken. Derivatives are a prerequisite. + #ifdef DRIVERS_ALLOW_DERIVATIVES + #define DRIVERS_ALLOW_FINE_DERIVATIVES + #endif + +// Dynamic looping: Requires an fp30 or newer profile. +// This makes phosphor mask resampling faster in some cases. Related errors: +// error C5013: profile does not support "for" statements and "for" could not +// be unrolled + //#define DRIVERS_ALLOW_DYNAMIC_BRANCHES + +// Without DRIVERS_ALLOW_DYNAMIC_BRANCHES, we need to use unrollable loops. +// Using one static loop avoids overhead if the user is right, but if the user +// is wrong (loops are allowed), breaking a loop into if-blocked pieces with a +// binary search can potentially save some iterations. However, it may fail: +// error C6001: Temporary register limit of 32 exceeded; 35 registers +// needed to compile program + //#define ACCOMODATE_POSSIBLE_DYNAMIC_LOOPS + +// tex2Dlod: Requires an fp40 or newer profile. This can be used to disable +// anisotropic filtering, thereby fixing related artifacts. Related errors: +// error C3004: function "float4 tex2Dlod(sampler2D, float4);" not supported in +// this profile + //#define DRIVERS_ALLOW_TEX2DLOD + +// tex2Dbias: Requires an fp30 or newer profile. This can be used to alleviate +// artifacts from anisotropic filtering and mipmapping. Related errors: +// error C3004: function "float4 tex2Dbias(sampler2D, float4);" not supported +// in this profile + //#define DRIVERS_ALLOW_TEX2DBIAS + +// Integrated graphics compatibility: Integrated graphics like Intel HD 4000 +// impose stricter limitations on register counts and instructions. Enable +// INTEGRATED_GRAPHICS_COMPATIBILITY_MODE if you still see error C6001 or: +// error C6002: Instruction limit of 1024 exceeded: 1523 instructions needed +// to compile program. +// Enabling integrated graphics compatibility mode will automatically disable: +// 1.) PHOSPHOR_MASK_MANUALLY_RESIZE: The phosphor mask will be softer. +// (This may be reenabled in a later release.) +// 2.) RUNTIME_GEOMETRY_MODE +// 3.) The high-quality 4x4 Gaussian resize for the bloom approximation + //#define INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + + +//////////////////////////// USER CODEPATH OPTIONS /////////////////////////// + +// To disable a #define option, turn its line into a comment with "//." + +// RUNTIME VS. COMPILE-TIME OPTIONS (Major Performance Implications): +// Enable runtime shader parameters in the Retroarch (etc.) GUI? They override +// many of the options in this file and allow real-time tuning, but many of +// them are slower. Disabling them and using this text file will boost FPS. +#define RUNTIME_SHADER_PARAMS_ENABLE +// Specify the phosphor bloom sigma at runtime? This option is 10% slower, but +// it's the only way to do a wide-enough full bloom with a runtime dot pitch. +#define RUNTIME_PHOSPHOR_BLOOM_SIGMA +// Specify antialiasing weight parameters at runtime? (Costs ~20% with cubics) +#define RUNTIME_ANTIALIAS_WEIGHTS +// Specify subpixel offsets at runtime? (WARNING: EXTREMELY EXPENSIVE!) +//#define RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS +// Make beam_horiz_filter and beam_horiz_linear_rgb_weight into runtime shader +// parameters? This will require more math or dynamic branching. +#define RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE +// Specify the tilt at runtime? This makes things about 3% slower. +#define RUNTIME_GEOMETRY_TILT +// Specify the geometry mode at runtime? +#define RUNTIME_GEOMETRY_MODE +// Specify the phosphor mask type (aperture grille, slot mask, shadow mask) and +// mode (Lanczos-resize, hardware resize, or tile 1:1) at runtime, even without +// dynamic branches? This is cheap if mask_resize_viewport_scale is small. +#define FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + +// PHOSPHOR MASK: +// Manually resize the phosphor mask for best results (slower)? Disabling this +// removes the option to do so, but it may be faster without dynamic branches. + #define PHOSPHOR_MASK_MANUALLY_RESIZE +// If we sinc-resize the mask, should we Lanczos-window it (slower but better)? + #define PHOSPHOR_MASK_RESIZE_LANCZOS_WINDOW +// Larger blurs are expensive, but we need them to blur larger triads. We can +// detect the right blur if the triad size is static or our profile allows +// dynamic branches, but otherwise we use the largest blur the user indicates +// they might need: + #define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_3_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_6_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_9_PIXELS + //#define PHOSPHOR_BLOOM_TRIADS_LARGER_THAN_12_PIXELS + // Here's a helpful chart: + // MaxTriadSize BlurSize MinTriadCountsByResolution + // 3.0 9.0 480/640/960/1920 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 6.0 17.0 240/320/480/960 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 9.0 25.0 160/213/320/640 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 12.0 31.0 120/160/240/480 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + // 18.0 43.0 80/107/160/320 triads at 1080p/1440p/2160p/4320p, 4:3 aspect + + +/////////////////////////////// USER PARAMETERS ////////////////////////////// + +// Note: Many of these static parameters are overridden by runtime shader +// parameters when those are enabled. However, many others are static codepath +// options that were cleaner or more convert to code as static constants. + +// GAMMA: + static const float crt_gamma_static = 2.5; // range [1, 5] + static const float lcd_gamma_static = 2.2; // range [1, 5] + +// LEVELS MANAGEMENT: + // Control the final multiplicative image contrast: + static const float levels_contrast_static = 1.0; // range [0, 4) + // We auto-dim to avoid clipping between passes and restore brightness + // later. Control the dim factor here: Lower values clip less but crush + // blacks more (static only for now). + static const float levels_autodim_temp = 0.5; // range (0, 1] default is 0.5 but that was unnecessarily dark for me, so I set it to 1.0 + +// HALATION/DIFFUSION/BLOOM: + // Halation weight: How much energy should be lost to electrons bounding + // around under the CRT glass and exciting random phosphors? + static const float halation_weight_static = 0.0; // range [0, 1] + // Refractive diffusion weight: How much light should spread/diffuse from + // refracting through the CRT glass? + static const float diffusion_weight_static = 0.075; // range [0, 1] + // Underestimate brightness: Bright areas bloom more, but we can base the + // bloom brightpass on a lower brightness to sharpen phosphors, or a higher + // brightness to soften them. Low values clip, but >= 0.8 looks okay. + static const float bloom_underestimate_levels_static = 0.8; // range [0, 5] + // Blur all colors more than necessary for a softer phosphor bloom? + static const float bloom_excess_static = 0.0; // range [0, 1] + // The BLOOM_APPROX pass approximates a phosphor blur early on with a small + // blurred resize of the input (convergence offsets are applied as well). + // There are three filter options (static option only for now): + // 0.) Bilinear resize: A fast, close approximation to a 4x4 resize + // if min_allowed_viewport_triads and the BLOOM_APPROX resolution are sane + // and beam_max_sigma is low. + // 1.) 3x3 resize blur: Medium speed, soft/smeared from bilinear blurring, + // always uses a static sigma regardless of beam_max_sigma or + // mask_num_triads_desired. + // 2.) True 4x4 Gaussian resize: Slowest, technically correct. + // These options are more pronounced for the fast, unbloomed shader version. +#ifndef RADEON_FIX + static const float bloom_approx_filter_static = 2.0; +#else + static const float bloom_approx_filter_static = 1.0; +#endif + +// ELECTRON BEAM SCANLINE DISTRIBUTION: + // How many scanlines should contribute light to each pixel? Using more + // scanlines is slower (especially for a generalized Gaussian) but less + // distorted with larger beam sigmas (especially for a pure Gaussian). The + // max_beam_sigma at which the closest unused weight is guaranteed < + // 1.0/255.0 (for a 3x antialiased pure Gaussian) is: + // 2 scanlines: max_beam_sigma = 0.2089; distortions begin ~0.34; 141.7 FPS pure, 131.9 FPS generalized + // 3 scanlines, max_beam_sigma = 0.3879; distortions begin ~0.52; 137.5 FPS pure; 123.8 FPS generalized + // 4 scanlines, max_beam_sigma = 0.5723; distortions begin ~0.70; 134.7 FPS pure; 117.2 FPS generalized + // 5 scanlines, max_beam_sigma = 0.7591; distortions begin ~0.89; 131.6 FPS pure; 112.1 FPS generalized + // 6 scanlines, max_beam_sigma = 0.9483; distortions begin ~1.08; 127.9 FPS pure; 105.6 FPS generalized + static const float beam_num_scanlines = 3.0; // range [2, 6] + // A generalized Gaussian beam varies shape with color too, now just width. + // It's slower but more flexible (static option only for now). + static const bool beam_generalized_gaussian = true; + // What kind of scanline antialiasing do you want? + // 0: Sample weights at 1x; 1: Sample weights at 3x; 2: Compute an integral + // Integrals are slow (especially for generalized Gaussians) and rarely any + // better than 3x antialiasing (static option only for now). + static const float beam_antialias_level = 1.0; // range [0, 2] + // Min/max standard deviations for scanline beams: Higher values widen and + // soften scanlines. Depending on other options, low min sigmas can alias. + static const float beam_min_sigma_static = 0.02; // range (0, 1] + static const float beam_max_sigma_static = 0.3; // range (0, 1] + // Beam width varies as a function of color: A power function (0) is more + // configurable, but a spherical function (1) gives the widest beam + // variability without aliasing (static option only for now). + static const float beam_spot_shape_function = 0.0; + // Spot shape power: Powers <= 1 give smoother spot shapes but lower + // sharpness. Powers >= 1.0 are awful unless mix/max sigmas are close. + static const float beam_spot_power_static = 1.0/3.0; // range (0, 16] + // Generalized Gaussian max shape parameters: Higher values give flatter + // scanline plateaus and steeper dropoffs, simultaneously widening and + // sharpening scanlines at the cost of aliasing. 2.0 is pure Gaussian, and + // values > ~40.0 cause artifacts with integrals. + static const float beam_min_shape_static = 2.0; // range [2, 32] + static const float beam_max_shape_static = 4.0; // range [2, 32] + // Generalized Gaussian shape power: Affects how quickly the distribution + // changes shape from Gaussian to steep/plateaued as color increases from 0 + // to 1.0. Higher powers appear softer for most colors, and lower powers + // appear sharper for most colors. + static const float beam_shape_power_static = 1.0/4.0; // range (0, 16] + // What filter should be used to sample scanlines horizontally? + // 0: Quilez (fast), 1: Gaussian (configurable), 2: Lanczos2 (sharp) + static const float beam_horiz_filter_static = 0.0; + // Standard deviation for horizontal Gaussian resampling: + static const float beam_horiz_sigma_static = 0.35; // range (0, 2/3] + // Do horizontal scanline sampling in linear RGB (correct light mixing), + // gamma-encoded RGB (darker, hard spot shape, may better match bandwidth- + // limiting circuitry in some CRT's), or a weighted avg.? + static const float beam_horiz_linear_rgb_weight_static = 1.0; // range [0, 1] + // Simulate scanline misconvergence? This needs 3x horizontal texture + // samples and 3x texture samples of BLOOM_APPROX and HALATION_BLUR in + // later passes (static option only for now). + static const bool beam_misconvergence = true; + // Convergence offsets in x/y directions for R/G/B scanline beams in units + // of scanlines. Positive offsets go right/down; ranges [-2, 2] + static const float2 convergence_offsets_r_static = float2(0.1, 0.2); + static const float2 convergence_offsets_g_static = float2(0.3, 0.4); + static const float2 convergence_offsets_b_static = float2(0.5, 0.6); + // Detect interlacing (static option only for now)? + static const bool interlace_detect = true; + // Assume 1080-line sources are interlaced? + static const bool interlace_1080i_static = false; + // For interlaced sources, assume TFF (top-field first) or BFF order? + // (Whether this matters depends on the nature of the interlaced input.) + static const bool interlace_bff_static = false; + +// ANTIALIASING: + // What AA level do you want for curvature/overscan/subpixels? Options: + // 0x (none), 1x (sample subpixels), 4x, 5x, 6x, 7x, 8x, 12x, 16x, 20x, 24x + // (Static option only for now) + static const float aa_level = 12.0; // range [0, 24] + // What antialiasing filter do you want (static option only)? Options: + // 0: Box (separable), 1: Box (cylindrical), + // 2: Tent (separable), 3: Tent (cylindrical), + // 4: Gaussian (separable), 5: Gaussian (cylindrical), + // 6: Cubic* (separable), 7: Cubic* (cylindrical, poor) + // 8: Lanczos Sinc (separable), 9: Lanczos Jinc (cylindrical, poor) + // * = Especially slow with RUNTIME_ANTIALIAS_WEIGHTS + static const float aa_filter = 6.0; // range [0, 9] + // Flip the sample grid on odd/even frames (static option only for now)? + static const bool aa_temporal = false; + // Use RGB subpixel offsets for antialiasing? The pixel is at green, and + // the blue offset is the negative r offset; range [0, 0.5] + static const float2 aa_subpixel_r_offset_static = float2(-1.0/3.0, 0.0);//float2(0.0); + // Cubics: See http://www.imagemagick.org/Usage/filter/#mitchell + // 1.) "Keys cubics" with B = 1 - 2C are considered the highest quality. + // 2.) C = 0.5 (default) is Catmull-Rom; higher C's apply sharpening. + // 3.) C = 1.0/3.0 is the Mitchell-Netravali filter. + // 4.) C = 0.0 is a soft spline filter. + static const float aa_cubic_c_static = 0.5; // range [0, 4] + // Standard deviation for Gaussian antialiasing: Try 0.5/aa_pixel_diameter. + static const float aa_gauss_sigma_static = 0.5; // range [0.0625, 1.0] + +// PHOSPHOR MASK: + // Mask type: 0 = aperture grille, 1 = slot mask, 2 = EDP shadow mask + static const float mask_type_static = 1.0; // range [0, 2] + // We can sample the mask three ways. Pick 2/3 from: Pretty/Fast/Flexible. + // 0.) Sinc-resize to the desired dot pitch manually (pretty/slow/flexible). + // This requires PHOSPHOR_MASK_MANUALLY_RESIZE to be #defined. + // 1.) Hardware-resize to the desired dot pitch (ugly/fast/flexible). This + // is halfway decent with LUT mipmapping but atrocious without it. + // 2.) Tile it without resizing at a 1:1 texel:pixel ratio for flat coords + // (pretty/fast/inflexible). Each input LUT has a fixed dot pitch. + // This mode reuses the same masks, so triads will be enormous unless + // you change the mask LUT filenames in your .cgp file. + static const float mask_sample_mode_static = 0.0; // range [0, 2] + // Prefer setting the triad size (0.0) or number on the screen (1.0)? + // If RUNTIME_PHOSPHOR_BLOOM_SIGMA isn't #defined, the specified triad size + // will always be used to calculate the full bloom sigma statically. + static const float mask_specify_num_triads_static = 0.0; // range [0, 1] + // Specify the phosphor triad size, in pixels. Each tile (usually with 8 + // triads) will be rounded to the nearest integer tile size and clamped to + // obey minimum size constraints (imposed to reduce downsize taps) and + // maximum size constraints (imposed to have a sane MASK_RESIZE FBO size). + // To increase the size limit, double the viewport-relative scales for the + // two MASK_RESIZE passes in crt-royale.cgp and user-cgp-contants.h. + // range [1, mask_texture_small_size/mask_triads_per_tile] + static const float mask_triad_size_desired_static = 24.0 / 8.0; + // If mask_specify_num_triads is 1.0/true, we'll go by this instead (the + // final size will be rounded and constrained as above); default 480.0 + static const float mask_num_triads_desired_static = 480.0; + // How many lobes should the sinc/Lanczos resizer use? More lobes require + // more samples and avoid moire a bit better, but some is unavoidable + // depending on the destination size (static option for now). + static const float mask_sinc_lobes = 3.0; // range [2, 4] + // The mask is resized using a variable number of taps in each dimension, + // but some Cg profiles always fetch a constant number of taps no matter + // what (no dynamic branching). We can limit the maximum number of taps if + // we statically limit the minimum phosphor triad size. Larger values are + // faster, but the limit IS enforced (static option only, forever); + // range [1, mask_texture_small_size/mask_triads_per_tile] + // TODO: Make this 1.0 and compensate with smarter sampling! + static const float mask_min_allowed_triad_size = 2.0; + +// GEOMETRY: + // Geometry mode: + // 0: Off (default), 1: Spherical mapping (like cgwg's), + // 2: Alt. spherical mapping (more bulbous), 3: Cylindrical/Trinitron + static const float geom_mode_static = 0.0; // range [0, 3] + // Radius of curvature: Measured in units of your viewport's diagonal size. + static const float geom_radius_static = 2.0; // range [1/(2*pi), 1024] + // View dist is the distance from the player to their physical screen, in + // units of the viewport's diagonal size. It controls the field of view. + static const float geom_view_dist_static = 2.0; // range [0.5, 1024] + // Tilt angle in radians (clockwise around up and right vectors): + static const float2 geom_tilt_angle_static = float2(0.0, 0.0); // range [-pi, pi] + // Aspect ratio: When the true viewport size is unknown, this value is used + // to help convert between the phosphor triad size and count, along with + // the mask_resize_viewport_scale constant from user-cgp-constants.h. Set + // this equal to Retroarch's display aspect ratio (DAR) for best results; + // range [1, geom_max_aspect_ratio from user-cgp-constants.h]; + // default (256/224)*(54/47) = 1.313069909 (see below) + static const float geom_aspect_ratio_static = 1.313069909; + // Before getting into overscan, here's some general aspect ratio info: + // - DAR = display aspect ratio = SAR * PAR; as in your Retroarch setting + // - SAR = storage aspect ratio = DAR / PAR; square pixel emulator frame AR + // - PAR = pixel aspect ratio = DAR / SAR; holds regardless of cropping + // Geometry processing has to "undo" the screen-space 2D DAR to calculate + // 3D view vectors, then reapplies the aspect ratio to the simulated CRT in + // uv-space. To ensure the source SAR is intended for a ~4:3 DAR, either: + // a.) Enable Retroarch's "Crop Overscan" + // b.) Readd horizontal padding: Set overscan to e.g. N*(1.0, 240.0/224.0) + // Real consoles use horizontal black padding in the signal, but emulators + // often crop this without cropping the vertical padding; a 256x224 [S]NES + // frame (8:7 SAR) is intended for a ~4:3 DAR, but a 256x240 frame is not. + // The correct [S]NES PAR is 54:47, found by blargg and NewRisingSun: + // http://board.zsnes.com/phpBB3/viewtopic.php?f=22&t=11928&start=50 + // http://forums.nesdev.com/viewtopic.php?p=24815#p24815 + // For flat output, it's okay to set DAR = [existing] SAR * [correct] PAR + // without doing a. or b., but horizontal image borders will be tighter + // than vertical ones, messing up curvature and overscan. Fixing the + // padding first corrects this. + // Overscan: Amount to "zoom in" before cropping. You can zoom uniformly + // or adjust x/y independently to e.g. readd horizontal padding, as noted + // above: Values < 1.0 zoom out; range (0, inf) + static const float2 geom_overscan_static = float2(1.0, 1.0);// * 1.005 * (1.0, 240/224.0) + // Compute a proper pixel-space to texture-space matrix even without ddx()/ + // ddy()? This is ~8.5% slower but improves antialiasing/subpixel filtering + // with strong curvature (static option only for now). + static const bool geom_force_correct_tangent_matrix = true; + +// BORDERS: + // Rounded border size in texture uv coords: + static const float border_size_static = 0.015; // range [0, 0.5] + // Border darkness: Moderate values darken the border smoothly, and high + // values make the image very dark just inside the border: + static const float border_darkness_static = 2.0; // range [0, inf) + // Border compression: High numbers compress border transitions, narrowing + // the dark border area. + static const float border_compress_static = 2.5; // range [1, inf) + + +#endif // USER_SETTINGS_H + +///////////////////////////// END USER-SETTINGS //////////////////////////// + +//#include "user-cgp-constants.h" + +///////////////////////// BEGIN USER-CGP-CONSTANTS ///////////////////////// + +#ifndef USER_CGP_CONSTANTS_H +#define USER_CGP_CONSTANTS_H + +// IMPORTANT: +// These constants MUST be set appropriately for the settings in crt-royale.cgp +// (or whatever related .cgp file you're using). If they aren't, you're likely +// to get artifacts, the wrong phosphor mask size, etc. I wish these could be +// set directly in the .cgp file to make things easier, but...they can't. + +// PASS SCALES AND RELATED CONSTANTS: +// Copy the absolute scale_x for BLOOM_APPROX. There are two major versions of +// this shader: One does a viewport-scale bloom, and the other skips it. The +// latter benefits from a higher bloom_approx_scale_x, so save both separately: +static const float bloom_approx_size_x = 320.0; +static const float bloom_approx_size_x_for_fake = 400.0; +// Copy the viewport-relative scales of the phosphor mask resize passes +// (MASK_RESIZE and the pass immediately preceding it): +static const float2 mask_resize_viewport_scale = float2(0.0625, 0.0625); +// Copy the geom_max_aspect_ratio used to calculate the MASK_RESIZE scales, etc.: +static const float geom_max_aspect_ratio = 4.0/3.0; + +// PHOSPHOR MASK TEXTURE CONSTANTS: +// Set the following constants to reflect the properties of the phosphor mask +// texture named in crt-royale.cgp. The shader optionally resizes a mask tile +// based on user settings, then repeats a single tile until filling the screen. +// The shader must know the input texture size (default 64x64), and to manually +// resize, it must also know the horizontal triads per tile (default 8). +static const float2 mask_texture_small_size = float2(64.0, 64.0); +static const float2 mask_texture_large_size = float2(512.0, 512.0); +static const float mask_triads_per_tile = 8.0; +// We need the average brightness of the phosphor mask to compensate for the +// dimming it causes. The following four values are roughly correct for the +// masks included with the shader. Update the value for any LUT texture you +// change. [Un]comment "#define PHOSPHOR_MASK_GRILLE14" depending on whether +// the loaded aperture grille uses 14-pixel or 15-pixel stripes (default 15). +//#define PHOSPHOR_MASK_GRILLE14 +static const float mask_grille14_avg_color = 50.6666666/255.0; + // TileableLinearApertureGrille14Wide7d33Spacing*.png + // TileableLinearApertureGrille14Wide10And6Spacing*.png +static const float mask_grille15_avg_color = 53.0/255.0; + // TileableLinearApertureGrille15Wide6d33Spacing*.png + // TileableLinearApertureGrille15Wide8And5d5Spacing*.png +static const float mask_slot_avg_color = 46.0/255.0; + // TileableLinearSlotMask15Wide9And4d5Horizontal8VerticalSpacing*.png + // TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing*.png +static const float mask_shadow_avg_color = 41.0/255.0; + // TileableLinearShadowMask*.png + // TileableLinearShadowMaskEDP*.png + +#ifdef PHOSPHOR_MASK_GRILLE14 + static const float mask_grille_avg_color = mask_grille14_avg_color; +#else + static const float mask_grille_avg_color = mask_grille15_avg_color; +#endif + + +#endif // USER_CGP_CONSTANTS_H + +////////////////////////// END USER-CGP-CONSTANTS ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +/////////////////////////////// FIXED SETTINGS /////////////////////////////// + +// Avoid dividing by zero; using a macro overloads for float, float2, etc.: +#define FIX_ZERO(c) (max(abs(c), 0.0000152587890625)) // 2^-16 + +// Ensure the first pass decodes CRT gamma and the last encodes LCD gamma. +#ifndef SIMULATE_CRT_ON_LCD + #define SIMULATE_CRT_ON_LCD +#endif + +// Manually tiling a manually resized texture creates texture coord derivative +// discontinuities and confuses anisotropic filtering, causing discolored tile +// seams in the phosphor mask. Workarounds: +// a.) Using tex2Dlod disables anisotropic filtering for tiled masks. It's +// downgraded to tex2Dbias without DRIVERS_ALLOW_TEX2DLOD #defined and +// disabled without DRIVERS_ALLOW_TEX2DBIAS #defined either. +// b.) "Tile flat twice" requires drawing two full tiles without border padding +// to the resized mask FBO, and it's incompatible with same-pass curvature. +// (Same-pass curvature isn't used but could be in the future...maybe.) +// c.) "Fix discontinuities" requires derivatives and drawing one tile with +// border padding to the resized mask FBO, but it works with same-pass +// curvature. It's disabled without DRIVERS_ALLOW_DERIVATIVES #defined. +// Precedence: a, then, b, then c (if multiple strategies are #defined). + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD // 129.7 FPS, 4x, flat; 101.8 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE // 128.1 FPS, 4x, flat; 101.5 at fullscreen + #define ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES // 124.4 FPS, 4x, flat; 97.4 at fullscreen +// Also, manually resampling the phosphor mask is slightly blurrier with +// anisotropic filtering. (Resampling with mipmapping is even worse: It +// creates artifacts, but only with the fully bloomed shader.) The difference +// is subtle with small triads, but you can fix it for a small cost. + //#define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + + +////////////////////////////// DERIVED SETTINGS ////////////////////////////// + +// Intel HD 4000 GPU's can't handle manual mask resizing (for now), setting the +// geometry mode at runtime, or a 4x4 true Gaussian resize. Disable +// incompatible settings ASAP. (INTEGRATED_GRAPHICS_COMPATIBILITY_MODE may be +// #defined by either user-settings.h or a wrapper .cg that #includes the +// current .cg pass.) +#ifdef INTEGRATED_GRAPHICS_COMPATIBILITY_MODE + #ifdef PHOSPHOR_MASK_MANUALLY_RESIZE + #undef PHOSPHOR_MASK_MANUALLY_RESIZE + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + // Mode 2 (4x4 Gaussian resize) won't work, and mode 1 (3x3 blur) is + // inferior in most cases, so replace 2.0 with 0.0: + static const float bloom_approx_filter = + bloom_approx_filter_static > 1.5 ? 0.0 : bloom_approx_filter_static; +#else + static const float bloom_approx_filter = bloom_approx_filter_static; +#endif + +// Disable slow runtime paths if static parameters are used. Most of these +// won't be a problem anyway once the params are disabled, but some will. +#ifndef RUNTIME_SHADER_PARAMS_ENABLE + #ifdef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #undef RUNTIME_PHOSPHOR_BLOOM_SIGMA + #endif + #ifdef RUNTIME_ANTIALIAS_WEIGHTS + #undef RUNTIME_ANTIALIAS_WEIGHTS + #endif + #ifdef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #undef RUNTIME_ANTIALIAS_SUBPIXEL_OFFSETS + #endif + #ifdef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #undef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #endif + #ifdef RUNTIME_GEOMETRY_TILT + #undef RUNTIME_GEOMETRY_TILT + #endif + #ifdef RUNTIME_GEOMETRY_MODE + #undef RUNTIME_GEOMETRY_MODE + #endif + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #undef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// Make tex2Dbias a backup for tex2Dlod for wider compatibility. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DBIAS +#endif +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #define ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS +#endif +// Rule out unavailable anisotropic compatibility strategies: +#ifndef DRIVERS_ALLOW_DERIVATIVES + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #endif + #ifdef ANTIALIAS_DISABLE_ANISOTROPIC + #undef ANTIALIAS_DISABLE_ANISOTROPIC + #endif +#endif +#ifndef DRIVERS_ALLOW_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif +// Prioritize anisotropic tiling compatibility strategies by performance and +// disable unused strategies. This concentrates all the nesting in one place. +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif +#else + #ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #undef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #endif + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #else + // ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE is only compatible with + // flat texture coords in the same pass, but that's all we use. + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + #ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #undef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + #endif + #endif + #endif +#endif +// The tex2Dlod and tex2Dbias strategies share a lot in common, and we can +// reduce some #ifdef nesting in the next section by essentially OR'ing them: +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DLOD + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +#ifdef ANISOTROPIC_TILING_COMPAT_TEX2DBIAS + #define ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY +#endif +// Prioritize anisotropic resampling compatibility strategies the same way: +#ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #undef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DBIAS + #endif +#endif + + +/////////////////////// DERIVED PHOSPHOR MASK CONSTANTS ////////////////////// + +// If we can use the large mipmapped LUT without mipmapping artifacts, we +// should: It gives us more options for using fewer samples. +#ifdef DRIVERS_ALLOW_TEX2DLOD + #ifdef ANISOTROPIC_RESAMPLING_COMPAT_TEX2DLOD + // TODO: Take advantage of this! + #define PHOSPHOR_MASK_RESIZE_MIPMAPPED_LUT + static const float2 mask_resize_src_lut_size = mask_texture_large_size; + #else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; + #endif +#else + static const float2 mask_resize_src_lut_size = mask_texture_small_size; +#endif + + +// tex2D's sampler2D parameter MUST be a uniform global, a uniform input to +// main_fragment, or a static alias of one of the above. This makes it hard +// to select the phosphor mask at runtime: We can't even assign to a uniform +// global in the vertex shader or select a sampler2D in the vertex shader and +// pass it to the fragment shader (even with explicit TEXUNIT# bindings), +// because it just gives us the input texture or a black screen. However, we +// can get around these limitations by calling tex2D three times with different +// uniform samplers (or resizing the phosphor mask three times altogether). +// With dynamic branches, we can process only one of these branches on top of +// quickly discarding fragments we don't need (cgc seems able to overcome +// limigations around dependent texture fetches inside of branches). Without +// dynamic branches, we have to process every branch for every fragment...which +// is slower. Runtime sampling mode selection is slower without dynamic +// branches as well. Let the user's static #defines decide if it's worth it. +#ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT +#else + #ifdef FORCE_RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #define RUNTIME_PHOSPHOR_MASK_MODE_TYPE_SELECT + #endif +#endif + +// We need to render some minimum number of tiles in the resize passes. +// We need at least 1.0 just to repeat a single tile, and we need extra +// padding beyond that for anisotropic filtering, discontinuitity fixing, +// antialiasing, same-pass curvature (not currently used), etc. First +// determine how many border texels and tiles we need, based on how the result +// will be sampled: +#ifdef GEOMETRY_EARLY + static const float max_subpixel_offset = aa_subpixel_r_offset_static.x; + // Most antialiasing filters have a base radius of 4.0 pixels: + static const float max_aa_base_pixel_border = 4.0 + + max_subpixel_offset; +#else + static const float max_aa_base_pixel_border = 0.0; +#endif +// Anisotropic filtering adds about 0.5 to the pixel border: +#ifndef ANISOTROPIC_TILING_COMPAT_TEX2DLOD_FAMILY + static const float max_aniso_pixel_border = max_aa_base_pixel_border + 0.5; +#else + static const float max_aniso_pixel_border = max_aa_base_pixel_border; +#endif +// Fixing discontinuities adds 1.0 more to the pixel border: +#ifdef ANISOTROPIC_TILING_COMPAT_FIX_DISCONTINUITIES + static const float max_tiled_pixel_border = max_aniso_pixel_border + 1.0; +#else + static const float max_tiled_pixel_border = max_aniso_pixel_border; +#endif +// Convert the pixel border to an integer texel border. Assume same-pass +// curvature about triples the texel frequency: +#ifdef GEOMETRY_EARLY + static const float max_mask_texel_border = + ceil(max_tiled_pixel_border * 3.0); +#else + static const float max_mask_texel_border = ceil(max_tiled_pixel_border); +#endif +// Convert the texel border to a tile border using worst-case assumptions: +static const float max_mask_tile_border = max_mask_texel_border/ + (mask_min_allowed_triad_size * mask_triads_per_tile); + +// Finally, set the number of resized tiles to render to MASK_RESIZE, and set +// the starting texel (inside borders) for sampling it. +#ifndef GEOMETRY_EARLY + #ifdef ANISOTROPIC_TILING_COMPAT_TILE_FLAT_TWICE + // Special case: Render two tiles without borders. Anisotropic + // filtering doesn't seem to be a problem here. + static const float mask_resize_num_tiles = 1.0 + 1.0; + static const float mask_start_texels = 0.0; + #else + static const float mask_resize_num_tiles = 1.0 + + 2.0 * max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; + #endif +#else + static const float mask_resize_num_tiles = 1.0 + 2.0*max_mask_tile_border; + static const float mask_start_texels = max_mask_texel_border; +#endif + +// We have to fit mask_resize_num_tiles into an FBO with a viewport scale of +// mask_resize_viewport_scale. This limits the maximum final triad size. +// Estimate the minimum number of triads we can split the screen into in each +// dimension (we'll be as correct as mask_resize_viewport_scale is): +static const float mask_resize_num_triads = + mask_resize_num_tiles * mask_triads_per_tile; +static const float2 min_allowed_viewport_triads = + float2(mask_resize_num_triads) / mask_resize_viewport_scale; + + +//////////////////////// COMMON MATHEMATICAL CONSTANTS /////////////////////// + +static const float pi = 3.141592653589; +// We often want to find the location of the previous texel, e.g.: +// const float2 curr_texel = uv * texture_size; +// const float2 prev_texel = floor(curr_texel - float2(0.5)) + float2(0.5); +// const float2 prev_texel_uv = prev_texel / texture_size; +// However, many GPU drivers round incorrectly around exact texel locations. +// We need to subtract a little less than 0.5 before flooring, and some GPU's +// require this value to be farther from 0.5 than others; define it here. +// const float2 prev_texel = +// floor(curr_texel - float2(under_half)) + float2(0.5); +static const float under_half = 0.4995; + + +#endif // DERIVED_SETTINGS_AND_CONSTANTS_H + +///////////////////////////// END DERIVED-SETTINGS-AND-CONSTANTS //////////////////////////// + +//#include "../../../../include/special-functions.h" + +/////////////////////////// BEGIN SPECIAL-FUNCTIONS ////////////////////////// + +#ifndef SPECIAL_FUNCTIONS_H +#define SPECIAL_FUNCTIONS_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file implements the following mathematical special functions: +// 1.) erf() = 2/sqrt(pi) * indefinite_integral(e**(-x**2)) +// 2.) gamma(s), a real-numbered extension of the integer factorial function +// It also implements normalized_ligamma(s, z), a normalized lower incomplete +// gamma function for s < 0.5 only. Both gamma() and normalized_ligamma() can +// be called with an _impl suffix to use an implementation version with a few +// extra precomputed parameters (which may be useful for the caller to reuse). +// See below for details. +// +// Design Rationale: +// Pretty much every line of code in this file is duplicated four times for +// different input types (float4/float3/float2/float). This is unfortunate, +// but Cg doesn't allow function templates. Macros would be far less verbose, +// but they would make the code harder to document and read. I don't expect +// these functions will require a whole lot of maintenance changes unless +// someone ever has need for more robust incomplete gamma functions, so code +// duplication seems to be the lesser evil in this case. + + +/////////////////////////// GAUSSIAN ERROR FUNCTION ////////////////////////// + +float4 erf6(float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Return an Abramowitz/Stegun approximation of erf(), where: + // erf(x) = 2/sqrt(pi) * integral(e**(-x**2)) + // This approximation has a max absolute error of 2.5*10**-5 + // with solid numerical robustness and efficiency. See: + // https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions + static const float4 one = float4(1.0); + const float4 sign_x = sign(x); + const float4 t = one/(one + 0.47047*abs(x)); + const float4 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float3 erf6(const float3 x) +{ + // Float3 version: + static const float3 one = float3(1.0); + const float3 sign_x = sign(x); + const float3 t = one/(one + 0.47047*abs(x)); + const float3 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float2 erf6(const float2 x) +{ + // Float2 version: + static const float2 one = float2(1.0); + const float2 sign_x = sign(x); + const float2 t = one/(one + 0.47047*abs(x)); + const float2 result = one - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float erf6(const float x) +{ + // Float version: + const float sign_x = sign(x); + const float t = 1.0/(1.0 + 0.47047*abs(x)); + const float result = 1.0 - t*(0.3480242 + t*(-0.0958798 + t*0.7478556))* + exp(-(x*x)); + return result * sign_x; +} + +float4 erft(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Approximate erf() with the hyperbolic tangent. The error is + // visually noticeable, but it's blazing fast and perceptually + // close...at least on ATI hardware. See: + // http://www.maplesoft.com/applications/view.aspx?SID=5525&view=html + // Warning: Only use this if your hardware drivers correctly implement + // tanh(): My nVidia 8800GTS returns garbage output. + return tanh(1.202760580 * x); +} + +float3 erft(const float3 x) +{ + // Float3 version: + return tanh(1.202760580 * x); +} + +float2 erft(const float2 x) +{ + // Float2 version: + return tanh(1.202760580 * x); +} + +float erft(const float x) +{ + // Float version: + return tanh(1.202760580 * x); +} + +inline float4 erf(const float4 x) +{ + // Requires: x is the standard parameter to erf(). + // Returns: Some approximation of erf(x), depending on user settings. + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float3 erf(const float3 x) +{ + // Float3 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float2 erf(const float2 x) +{ + // Float2 version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + +inline float erf(const float x) +{ + // Float version: + #ifdef ERF_FAST_APPROXIMATION + return erft(x); + #else + return erf6(x); + #endif +} + + +/////////////////////////// COMPLETE GAMMA FUNCTION ////////////////////////// + +float4 gamma_impl(const float4 s, const float4 s_inv) +{ + // Requires: 1.) s is the standard parameter to the gamma function, and + // it should lie in the [0, 36] range. + // 2.) s_inv = 1.0/s. This implementation function requires + // the caller to precompute this value, giving users the + // opportunity to reuse it. + // Returns: Return approximate gamma function (real-numbered factorial) + // output using the Lanczos approximation with two coefficients + // calculated using Paul Godfrey's method here: + // http://my.fit.edu/~gabdo/gamma.txt + // An optimal g value for s in [0, 36] is ~1.12906830989, with + // a maximum relative error of 0.000463 for 2**16 equally + // evals. We could use three coeffs (0.0000346 error) without + // hurting latency, but this allows more parallelism with + // outside instructions. + static const float4 g = float4(1.12906830989); + static const float4 c0 = float4(0.8109119309638332633713423362694399653724431); + static const float4 c1 = float4(0.4808354605142681877121661197951496120000040); + static const float4 e = float4(2.71828182845904523536028747135266249775724709); + const float4 sph = s + float4(0.5); + const float4 lanczos_sum = c0 + c1/(s + float4(1.0)); + const float4 base = (sph + g)/e; // or (s + g + float4(0.5))/e + // gamma(s + 1) = base**sph * lanczos_sum; divide by s for gamma(s). + // This has less error for small s's than (s -= 1.0) at the beginning. + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float3 gamma_impl(const float3 s, const float3 s_inv) +{ + // Float3 version: + static const float3 g = float3(1.12906830989); + static const float3 c0 = float3(0.8109119309638332633713423362694399653724431); + static const float3 c1 = float3(0.4808354605142681877121661197951496120000040); + static const float3 e = float3(2.71828182845904523536028747135266249775724709); + const float3 sph = s + float3(0.5); + const float3 lanczos_sum = c0 + c1/(s + float3(1.0)); + const float3 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float2 gamma_impl(const float2 s, const float2 s_inv) +{ + // Float2 version: + static const float2 g = float2(1.12906830989); + static const float2 c0 = float2(0.8109119309638332633713423362694399653724431); + static const float2 c1 = float2(0.4808354605142681877121661197951496120000040); + static const float2 e = float2(2.71828182845904523536028747135266249775724709); + const float2 sph = s + float2(0.5); + const float2 lanczos_sum = c0 + c1/(s + float2(1.0)); + const float2 base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float gamma_impl(const float s, const float s_inv) +{ + // Float version: + static const float g = 1.12906830989; + static const float c0 = 0.8109119309638332633713423362694399653724431; + static const float c1 = 0.4808354605142681877121661197951496120000040; + static const float e = 2.71828182845904523536028747135266249775724709; + const float sph = s + 0.5; + const float lanczos_sum = c0 + c1/(s + 1.0); + const float base = (sph + g)/e; + return (pow(base, sph) * lanczos_sum) * s_inv; +} + +float4 gamma(const float4 s) +{ + // Requires: s is the standard parameter to the gamma function, and it + // should lie in the [0, 36] range. + // Returns: Return approximate gamma function output with a maximum + // relative error of 0.000463. See gamma_impl for details. + return gamma_impl(s, float4(1.0)/s); +} + +float3 gamma(const float3 s) +{ + // Float3 version: + return gamma_impl(s, float3(1.0)/s); +} + +float2 gamma(const float2 s) +{ + // Float2 version: + return gamma_impl(s, float2(1.0)/s); +} + +float gamma(const float s) +{ + // Float version: + return gamma_impl(s, 1.0/s); +} + + +//////////////// INCOMPLETE GAMMA FUNCTIONS (RESTRICTED INPUT) /////////////// + +// Lower incomplete gamma function for small s and z (implementation): +float4 ligamma_small_z_impl(const float4 s, const float4 z, const float4 s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) z <= ~0.775075 + // 3.) s_inv = 1.0/s (precomputed for outside reuse) + // Returns: A series representation for the lower incomplete gamma + // function for small s and small z (4 terms). + // The actual "rolled up" summation looks like: + // last_sign = 1.0; last_pow = 1.0; last_factorial = 1.0; + // sum = last_sign * last_pow / ((s + k) * last_factorial) + // for(int i = 0; i < 4; ++i) + // { + // last_sign *= -1.0; last_pow *= z; last_factorial *= i; + // sum += last_sign * last_pow / ((s + k) * last_factorial); + // } + // Unrolled, constant-unfolded and arranged for madds and parallelism: + const float4 scale = pow(z, s); + float4 sum = s_inv; // Summation iteration 0 result + // Summation iterations 1, 2, and 3: + const float4 z_sq = z*z; + const float4 denom1 = s + float4(1.0); + const float4 denom2 = 2.0*s + float4(4.0); + const float4 denom3 = 6.0*s + float4(18.0); + //float4 denom4 = 24.0*s + float4(96.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + //sum += z_sq * z_sq / denom4; + // Scale and return: + return scale * sum; +} + +float3 ligamma_small_z_impl(const float3 s, const float3 z, const float3 s_inv) +{ + // Float3 version: + const float3 scale = pow(z, s); + float3 sum = s_inv; + const float3 z_sq = z*z; + const float3 denom1 = s + float3(1.0); + const float3 denom2 = 2.0*s + float3(4.0); + const float3 denom3 = 6.0*s + float3(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float2 ligamma_small_z_impl(const float2 s, const float2 z, const float2 s_inv) +{ + // Float2 version: + const float2 scale = pow(z, s); + float2 sum = s_inv; + const float2 z_sq = z*z; + const float2 denom1 = s + float2(1.0); + const float2 denom2 = 2.0*s + float2(4.0); + const float2 denom3 = 6.0*s + float2(18.0); + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +float ligamma_small_z_impl(const float s, const float z, const float s_inv) +{ + // Float version: + const float scale = pow(z, s); + float sum = s_inv; + const float z_sq = z*z; + const float denom1 = s + 1.0; + const float denom2 = 2.0*s + 4.0; + const float denom3 = 6.0*s + 18.0; + sum -= z/denom1; + sum += z_sq/denom2; + sum -= z * z_sq/denom3; + return scale * sum; +} + +// Upper incomplete gamma function for small s and large z (implementation): +float4 uigamma_large_z_impl(const float4 s, const float4 z) +{ + // Requires: 1.) s < ~0.5 + // 2.) z > ~0.775075 + // Returns: Gauss's continued fraction representation for the upper + // incomplete gamma function (4 terms). + // The "rolled up" continued fraction looks like this. The denominator + // is truncated, and it's calculated "from the bottom up:" + // denom = float4('inf'); + // float4 one = float4(1.0); + // for(int i = 4; i > 0; --i) + // { + // denom = ((i * 2.0) - one) + z - s + (i * (s - i))/denom; + // } + // Unrolled and constant-unfolded for madds and parallelism: + const float4 numerator = pow(z, s) * exp(-z); + float4 denom = float4(7.0) + z - s; + denom = float4(5.0) + z - s + (3.0*s - float4(9.0))/denom; + denom = float4(3.0) + z - s + (2.0*s - float4(4.0))/denom; + denom = float4(1.0) + z - s + (s - float4(1.0))/denom; + return numerator / denom; +} + +float3 uigamma_large_z_impl(const float3 s, const float3 z) +{ + // Float3 version: + const float3 numerator = pow(z, s) * exp(-z); + float3 denom = float3(7.0) + z - s; + denom = float3(5.0) + z - s + (3.0*s - float3(9.0))/denom; + denom = float3(3.0) + z - s + (2.0*s - float3(4.0))/denom; + denom = float3(1.0) + z - s + (s - float3(1.0))/denom; + return numerator / denom; +} + +float2 uigamma_large_z_impl(const float2 s, const float2 z) +{ + // Float2 version: + const float2 numerator = pow(z, s) * exp(-z); + float2 denom = float2(7.0) + z - s; + denom = float2(5.0) + z - s + (3.0*s - float2(9.0))/denom; + denom = float2(3.0) + z - s + (2.0*s - float2(4.0))/denom; + denom = float2(1.0) + z - s + (s - float2(1.0))/denom; + return numerator / denom; +} + +float uigamma_large_z_impl(const float s, const float z) +{ + // Float version: + const float numerator = pow(z, s) * exp(-z); + float denom = 7.0 + z - s; + denom = 5.0 + z - s + (3.0*s - 9.0)/denom; + denom = 3.0 + z - s + (2.0*s - 4.0)/denom; + denom = 1.0 + z - s + (s - 1.0)/denom; + return numerator / denom; +} + +// Normalized lower incomplete gamma function for small s (implementation): +float4 normalized_ligamma_impl(const float4 s, const float4 z, + const float4 s_inv, const float4 gamma_s_inv) +{ + // Requires: 1.) s < ~0.5 + // 2.) s_inv = 1/s (precomputed for outside reuse) + // 3.) gamma_s_inv = 1/gamma(s) (precomputed for outside reuse) + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. Since we only care about s < 0.5, we only need + // to evaluate two branches (not four) based on z. Each branch + // uses four terms, with a max relative error of ~0.00182. The + // branch threshold and specifics were adapted for fewer terms + // from Gil/Segura/Temme's paper here: + // http://oai.cwi.nl/oai/asset/20433/20433B.pdf + // Evaluate both branches: Real branches test slower even when available. + static const float4 thresh = float4(0.775075); + bool4 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + z_is_large.w = z.w > thresh.w; + const float4 large_z = float4(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float4 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + // Combine the results from both branches: + bool4 inverse_z_is_large = not(z_is_large); + return large_z * float4(z_is_large) + small_z * float4(inverse_z_is_large); +} + +float3 normalized_ligamma_impl(const float3 s, const float3 z, + const float3 s_inv, const float3 gamma_s_inv) +{ + // Float3 version: + static const float3 thresh = float3(0.775075); + bool3 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + z_is_large.z = z.z > thresh.z; + const float3 large_z = float3(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float3 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool3 inverse_z_is_large = not(z_is_large); + return large_z * float3(z_is_large) + small_z * float3(inverse_z_is_large); +} + +float2 normalized_ligamma_impl(const float2 s, const float2 z, + const float2 s_inv, const float2 gamma_s_inv) +{ + // Float2 version: + static const float2 thresh = float2(0.775075); + bool2 z_is_large; + z_is_large.x = z.x > thresh.x; + z_is_large.y = z.y > thresh.y; + const float2 large_z = float2(1.0) - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float2 small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + bool2 inverse_z_is_large = not(z_is_large); + return large_z * float2(z_is_large) + small_z * float2(inverse_z_is_large); +} + +float normalized_ligamma_impl(const float s, const float z, + const float s_inv, const float gamma_s_inv) +{ + // Float version: + static const float thresh = 0.775075; + const bool z_is_large = z > thresh; + const float large_z = 1.0 - uigamma_large_z_impl(s, z) * gamma_s_inv; + const float small_z = ligamma_small_z_impl(s, z, s_inv) * gamma_s_inv; + return large_z * float(z_is_large) + small_z * float(!z_is_large); +} + +// Normalized lower incomplete gamma function for small s: +float4 normalized_ligamma(const float4 s, const float4 z) +{ + // Requires: s < ~0.5 + // Returns: Approximate the normalized lower incomplete gamma function + // for s < 0.5. See normalized_ligamma_impl() for details. + const float4 s_inv = float4(1.0)/s; + const float4 gamma_s_inv = float4(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float3 normalized_ligamma(const float3 s, const float3 z) +{ + // Float3 version: + const float3 s_inv = float3(1.0)/s; + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float2 normalized_ligamma(const float2 s, const float2 z) +{ + // Float2 version: + const float2 s_inv = float2(1.0)/s; + const float2 gamma_s_inv = float2(1.0)/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +float normalized_ligamma(const float s, const float z) +{ + // Float version: + const float s_inv = 1.0/s; + const float gamma_s_inv = 1.0/gamma_impl(s, s_inv); + return normalized_ligamma_impl(s, z, s_inv, gamma_s_inv); +} + +#endif // SPECIAL_FUNCTIONS_H + +//////////////////////////// END SPECIAL-FUNCTIONS /////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(texture(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +//////////////////////////////// END INCLUDES //////////////////////////////// + +///////////////////////////// SCANLINE FUNCTIONS ///////////////////////////// + +inline float3 get_gaussian_sigma(const float3 color, const float sigma_range) +{ + // Requires: Globals: + // 1.) beam_min_sigma and beam_max_sigma are global floats + // containing the desired minimum and maximum beam standard + // deviations, for dim and bright colors respectively. + // 2.) beam_max_sigma must be > 0.0 + // 3.) beam_min_sigma must be in (0.0, beam_max_sigma] + // 4.) beam_spot_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) sigma_range = beam_max_sigma - beam_min_sigma; we take + // sigma_range as a parameter to avoid repeated computation + // when beam_{min, max}_sigma are runtime shader parameters + // Optional: Users may set beam_spot_shape_function to 1 to define the + // inner f(color) subfunction (see below) as: + // f(color) = sqrt(1.0 - (color - 1.0)*(color - 1.0)) + // Otherwise (technically, if beam_spot_shape_function < 0.5): + // f(color) = pow(color, beam_spot_power) + // Returns: The standard deviation of the Gaussian beam for "color:" + // sigma = beam_min_sigma + sigma_range * f(color) + // Details/Discussion: + // The beam's spot shape vaguely resembles an aspect-corrected f() in the + // range [0, 1] (not quite, but it's related). f(color) = color makes + // spots look like diamonds, and a spherical function or cube balances + // between variable width and a soft/realistic shape. A beam_spot_power + // > 1.0 can produce an ugly spot shape and more initial clipping, but the + // final shape also differs based on the horizontal resampling filter and + // the phosphor bloom. For instance, resampling horizontally in nonlinear + // light and/or with a sharp (e.g. Lanczos) filter will sharpen the spot + // shape, but a sixth root is still quite soft. A power function (default + // 1.0/3.0 beam_spot_power) is most flexible, but a fixed spherical curve + // has the highest variability without an awful spot shape. + // + // beam_min_sigma affects scanline sharpness/aliasing in dim areas, and its + // difference from beam_max_sigma affects beam width variability. It only + // affects clipping [for pure Gaussians] if beam_spot_power > 1.0 (which is + // a conservative estimate for a more complex constraint). + // + // beam_max_sigma affects clipping and increasing scanline width/softness + // as color increases. The wider this is, the more scanlines need to be + // evaluated to avoid distortion. For a pure Gaussian, the max_beam_sigma + // at which the first unused scanline always has a weight < 1.0/255.0 is: + // num scanlines = 2, max_beam_sigma = 0.2089; distortions begin ~0.34 + // num scanlines = 3, max_beam_sigma = 0.3879; distortions begin ~0.52 + // num scanlines = 4, max_beam_sigma = 0.5723; distortions begin ~0.70 + // num scanlines = 5, max_beam_sigma = 0.7591; distortions begin ~0.89 + // num scanlines = 6, max_beam_sigma = 0.9483; distortions begin ~1.08 + // Generalized Gaussians permit more leeway here as steepness increases. + if(beam_spot_shape_function < 0.5) + { + // Use a power function: + return float3(beam_min_sigma) + sigma_range * + pow(color, float3(beam_spot_power)); + } + else + { + // Use a spherical function: + const float3 color_minus_1 = color - float3(1.0); + return float3(beam_min_sigma) + sigma_range * + sqrt(float3(1.0) - color_minus_1*color_minus_1); + } +} + +inline float3 get_generalized_gaussian_beta(const float3 color, + const float shape_range) +{ + // Requires: Globals: + // 1.) beam_min_shape and beam_max_shape are global floats + // containing the desired min/max generalized Gaussian + // beta parameters, for dim and bright colors respectively. + // 2.) beam_max_shape must be >= 2.0 + // 3.) beam_min_shape must be in [2.0, beam_max_shape] + // 4.) beam_shape_power must be defined as a global float. + // Parameters: + // 1.) color is the underlying source color along a scanline + // 2.) shape_range = beam_max_shape - beam_min_shape; we take + // shape_range as a parameter to avoid repeated computation + // when beam_{min, max}_shape are runtime shader parameters + // Returns: The type-I generalized Gaussian "shape" parameter beta for + // the given color. + // Details/Discussion: + // Beta affects the scanline distribution as follows: + // a.) beta < 2.0 narrows the peak to a spike with a discontinuous slope + // b.) beta == 2.0 just degenerates to a Gaussian + // c.) beta > 2.0 flattens and widens the peak, then drops off more steeply + // than a Gaussian. Whereas high sigmas widen and soften peaks, high + // beta widen and sharpen peaks at the risk of aliasing. + // Unlike high beam_spot_powers, high beam_shape_powers actually soften shape + // transitions, whereas lower ones sharpen them (at the risk of aliasing). + return beam_min_shape + shape_range * pow(color, float3(beam_shape_power)); +} + +float3 scanline_gaussian_integral_contrib(const float3 dist, + const float3 color, const float pixel_height, const float sigma_range) +{ + // Requires: 1.) dist is the distance of the [potentially separate R/G/B] + // point(s) from a scanline in units of scanlines, where + // 1.0 means the sample point straddles the next scanline. + // 2.) color is the underlying source color along a scanline. + // 3.) pixel_height is the output pixel height in scanlines. + // 4.) Requirements of get_gaussian_sigma() must be met. + // Returns: Return a scanline's light output over a given pixel. + // Details: + // The CRT beam profile follows a roughly Gaussian distribution which is + // wider for bright colors than dark ones. The integral over the full + // range of a Gaussian function is always 1.0, so we can vary the beam + // with a standard deviation without affecting brightness. 'x' = distance: + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + // gaussian integral = 0.5 (1.0 + erf(x/(sigma * sqrt(2)))) + // Use a numerical approximation of the "error function" (the Gaussian + // indefinite integral) to find the definite integral of the scanline's + // average brightness over a given pixel area. Even if curved coords were + // used in this pass, a flat scalar pixel height works almost as well as a + // pixel height computed from a full pixel-space to scanline-space matrix. + const float3 sigma = get_gaussian_sigma(color, sigma_range); + const float3 ph_offset = float3(pixel_height * 0.5); + const float3 denom_inv = 1.0/(sigma*sqrt(2.0)); + const float3 integral_high = erf((dist + ph_offset)*denom_inv); + const float3 integral_low = erf((dist - ph_offset)*denom_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_generalized_gaussian_integral_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel. + // A generalized Gaussian distribution allows the shape (beta) to vary + // as well as the width (alpha). "gamma" refers to the gamma function: + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + // ligamma(s, z) is the lower incomplete gamma function, for which we only + // implement two of four branches (because we keep 1/beta <= 0.5): + // generalized integral = 0.5 + 0.5* sign(x) * + // ligamma(1/beta, (|x|/alpha)**beta)/gamma(1/beta) + // See get_generalized_gaussian_beta() for a discussion of beta. + // We base alpha on the intended Gaussian sigma, but it only strictly + // models models standard deviation at beta == 2, because the standard + // deviation depends on both alpha and beta (keeping alpha independent is + // faster and preserves intuitive behavior and a full spectrum of results). + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + const float3 alpha_inv = float3(1.0)/alpha; + const float3 s = float3(1.0)/beta; + const float3 ph_offset = float3(pixel_height * 0.5); + // Pass beta to gamma_impl to avoid repeated divides. Similarly pass + // beta (i.e. 1/s) and 1/gamma(s) to normalized_ligamma_impl. + const float3 gamma_s_inv = float3(1.0)/gamma_impl(s, beta); + const float3 dist1 = dist + ph_offset; + const float3 dist0 = dist - ph_offset; + const float3 integral_high = sign(dist1) * normalized_ligamma_impl( + s, pow(abs(dist1)*alpha_inv, beta), beta, gamma_s_inv); + const float3 integral_low = sign(dist0) * normalized_ligamma_impl( + s, pow(abs(dist0)*alpha_inv, beta), beta, gamma_s_inv); + return color * 0.5*(integral_high - integral_low)/pixel_height; +} + +float3 scanline_gaussian_sampled_contrib(const float3 dist, const float3 color, + const float pixel_height, const float sigma_range) +{ + // See scanline_gaussian integral_contrib() for detailed comments! + // gaussian sample = 1/(sigma*sqrt(2*pi)) * e**(-(x**2)/(2*sigma**2)) + const float3 sigma = get_gaussian_sigma(color, sigma_range); + // Avoid repeated divides: + const float3 sigma_inv = float3(1.0)/sigma; + const float3 inner_denom_inv = 0.5 * sigma_inv * sigma_inv; + const float3 outer_denom_inv = sigma_inv/sqrt(2.0*pi); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel away in each direction as well: + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three pure Gaussian samples: + const float3 scale = color/3.0 * outer_denom_inv; + const float3 weight1 = exp(-(dist*dist)*inner_denom_inv); + const float3 weight2 = exp(-(dist2*dist2)*inner_denom_inv); + const float3 weight3 = exp(-(dist3*dist3)*inner_denom_inv); + return scale * (weight1 + weight2 + weight3); + } + else + { + return color*exp(-(dist*dist)*inner_denom_inv)*outer_denom_inv; + } +} + +float3 scanline_generalized_gaussian_sampled_contrib(float3 dist, + float3 color, float pixel_height, float sigma_range, + float shape_range) +{ + // See scanline_generalized_gaussian_integral_contrib() for details! + // generalized sample = + // beta/(2*alpha*gamma(1/beta)) * e**(-(|x|/alpha)**beta) + const float3 alpha = sqrt(2.0) * get_gaussian_sigma(color, sigma_range); + const float3 beta = get_generalized_gaussian_beta(color, shape_range); + // Avoid repeated divides: + const float3 alpha_inv = float3(1.0)/alpha; + const float3 beta_inv = float3(1.0)/beta; + const float3 scale = color * beta * 0.5 * alpha_inv / + gamma_impl(beta_inv, beta); + if(beam_antialias_level > 0.5) + { + // Sample 1/3 pixel closer to and farther from the scanline too. + const float3 sample_offset = float3(pixel_height/3.0); + const float3 dist2 = dist + sample_offset; + const float3 dist3 = abs(dist - sample_offset); + // Average three generalized Gaussian samples: + const float3 weight1 = exp(-pow(abs(dist*alpha_inv), beta)); + const float3 weight2 = exp(-pow(abs(dist2*alpha_inv), beta)); + const float3 weight3 = exp(-pow(abs(dist3*alpha_inv), beta)); + return scale/3.0 * (weight1 + weight2 + weight3); + } + else + { + return scale * exp(-pow(abs(dist*alpha_inv), beta)); + } +} + +inline float3 scanline_contrib(float3 dist, float3 color, + float pixel_height, const float sigma_range, const float shape_range) +{ + // Requires: 1.) Requirements of scanline_gaussian_integral_contrib() + // must be met. + // 2.) Requirements of get_gaussian_sigma() must be met. + // 3.) Requirements of get_generalized_gaussian_beta() must be + // met. + // Returns: Return a scanline's light output over a given pixel, using + // a generalized or pure Gaussian distribution and sampling or + // integrals as desired by user codepath choices. + if(beam_generalized_gaussian) + { + if(beam_antialias_level > 1.5) + { + return scanline_generalized_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + else + { + return scanline_generalized_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range, shape_range); + } + } + else + { + if(beam_antialias_level > 1.5) + { + return scanline_gaussian_integral_contrib( + dist, color, pixel_height, sigma_range); + } + else + { + return scanline_gaussian_sampled_contrib( + dist, color, pixel_height, sigma_range); + } + } +} + +inline float3 get_raw_interpolated_color(const float3 color0, + const float3 color1, const float3 color2, const float3 color3, + const float4 weights) +{ + // Use max to avoid bizarre artifacts from negative colors: + return max(mul(weights, float4x3(color0, color1, color2, color3)), 0.0); +} + +float3 get_interpolated_linear_color(const float3 color0, const float3 color1, + const float3 color2, const float3 color3, const float4 weights) +{ + // Requires: 1.) Requirements of include/gamma-management.h must be met: + // intermediate_gamma must be globally defined, and input + // colors are interpreted as linear RGB unless you #define + // GAMMA_ENCODE_EVERY_FBO (in which case they are + // interpreted as gamma-encoded with intermediate_gamma). + // 2.) color0-3 are colors sampled from a texture with tex2D(). + // They are interpreted as defined in requirement 1. + // 3.) weights contains weights for each color, summing to 1.0. + // 4.) beam_horiz_linear_rgb_weight must be defined as a global + // float in [0.0, 1.0] describing how much blending should + // be done in linear RGB (rest is gamma-corrected RGB). + // 5.) RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE must be #defined + // if beam_horiz_linear_rgb_weight is anything other than a + // static constant, or we may try branching at runtime + // without dynamic branches allowed (slow). + // Returns: Return an interpolated color lookup between the four input + // colors based on the weights in weights. The final color will + // be a linear RGB value, but the blending will be done as + // indicated above. + const float intermediate_gamma = get_intermediate_gamma(); + // Branch if beam_horiz_linear_rgb_weight is static (for free) or if the + // profile allows dynamic branches (faster than computing extra pows): + #ifndef RUNTIME_SCANLINES_HORIZ_FILTER_COLORSPACE + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #else + #ifdef DRIVERS_ALLOW_DYNAMIC_BRANCHES + #define SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + #endif + #endif + #ifdef SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT + // beam_horiz_linear_rgb_weight is static, so we can branch: + #ifdef GAMMA_ENCODE_EVERY_FBO + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), float3(intermediate_gamma)); + if(beam_horiz_linear_rgb_weight > 0.0) + { + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return gamma_mixed_color; + } + #else + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + if(beam_horiz_linear_rgb_weight < 1.0) + { + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + } + else + { + return linear_mixed_color; + } + #endif // GAMMA_ENCODE_EVERY_FBO + #else + #ifdef GAMMA_ENCODE_EVERY_FBO + // Inputs: color0-3 are colors in gamma-encoded RGB. + const float3 gamma_mixed_color = pow(get_raw_interpolated_color( + color0, color1, color2, color3, weights), intermediate_gamma); + const float3 linear_mixed_color = get_raw_interpolated_color( + pow(color0, float3(intermediate_gamma)), + pow(color1, float3(intermediate_gamma)), + pow(color2, float3(intermediate_gamma)), + pow(color3, float3(intermediate_gamma)), + weights); + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #else + // Inputs: color0-3 are colors in linear RGB. + const float3 linear_mixed_color = get_raw_interpolated_color( + color0, color1, color2, color3, weights); + const float3 gamma_mixed_color = get_raw_interpolated_color( + pow(color0, float3(1.0/intermediate_gamma)), + pow(color1, float3(1.0/intermediate_gamma)), + pow(color2, float3(1.0/intermediate_gamma)), + pow(color3, float3(1.0/intermediate_gamma)), + weights); + // wtf fixme +// const float beam_horiz_linear_rgb_weight1 = 1.0; + return lerp(gamma_mixed_color, linear_mixed_color, + beam_horiz_linear_rgb_weight); + #endif // GAMMA_ENCODE_EVERY_FBO + #endif // SCANLINES_BRANCH_FOR_LINEAR_RGB_WEIGHT +} + +float3 get_scanline_color(const sampler2D tex, const float2 scanline_uv, + const float2 uv_step_x, const float4 weights) +{ + // Requires: 1.) scanline_uv must be vertically snapped to the caller's + // desired line or scanline and horizontally snapped to the + // texel just left of the output pixel (color1) + // 2.) uv_step_x must contain the horizontal uv distance + // between texels. + // 3.) weights must contain interpolation filter weights for + // color0, color1, color2, and color3, where color1 is just + // left of the output pixel. + // Returns: Return a horizontally interpolated texture lookup using 2-4 + // nearby texels, according to weights and the conventions of + // get_interpolated_linear_color(). + // We can ignore the outside texture lookups for Quilez resampling. + const float3 color1 = COMPAT_TEXTURE(tex, scanline_uv).rgb; + const float3 color2 = COMPAT_TEXTURE(tex, scanline_uv + uv_step_x).rgb; + float3 color0 = float3(0.0); + float3 color3 = float3(0.0); + if(beam_horiz_filter > 0.5) + { + color0 = COMPAT_TEXTURE(tex, scanline_uv - uv_step_x).rgb; + color3 = COMPAT_TEXTURE(tex, scanline_uv + 2.0 * uv_step_x).rgb; + } + // Sample the texture as-is, whether it's linear or gamma-encoded: + // get_interpolated_linear_color() will handle the difference. + return get_interpolated_linear_color(color0, color1, color2, color3, weights); +} + +float3 sample_single_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Snap to the previous texel and get sample dists from 2/4 nearby texels: + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel = + floor(curr_texel - float2(under_half)) + float2(0.5); + const float2 prev_texel_hor = float2(prev_texel.x, curr_texel.y); + const float2 prev_texel_hor_uv = prev_texel_hor * texture_size_inv; + const float prev_dist = curr_texel.x - prev_texel_hor.x; + const float4 sample_dists = float4(1.0 + prev_dist, prev_dist, + 1.0 - prev_dist, 2.0 - prev_dist); + // Get Quilez, Lanczos2, or Gaussian resize weights for 2/4 nearby texels: + float4 weights; + if(beam_horiz_filter < 0.5) + { + // Quilez: + const float x = sample_dists.y; + const float w2 = x*x*x*(x*(x*6.0 - 15.0) + 10.0); + weights = float4(0.0, 1.0 - w2, w2, 0.0); + } + else if(beam_horiz_filter < 1.5) + { + // Gaussian: + float inner_denom_inv = 1.0/(2.0*beam_horiz_sigma*beam_horiz_sigma); + weights = exp(-(sample_dists*sample_dists)*inner_denom_inv); + } + else + { + // Lanczos2: + const float4 pi_dists = FIX_ZERO(sample_dists * pi); + weights = 2.0 * sin(pi_dists) * sin(pi_dists * 0.5) / + (pi_dists * pi_dists); + } + // Ensure the weight sum == 1.0: + const float4 final_weights = weights/dot(weights, float4(1.0)); + // Get the interpolated horizontal scanline color: + const float2 uv_step_x = float2(texture_size_inv.x, 0.0); + return get_scanline_color( + tex, prev_texel_hor_uv, uv_step_x, final_weights); +} + +float3 sample_rgb_scanline_horizontal(const sampler2D tex, + const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv) +{ + // TODO: Add function requirements. + // Rely on a helper to make convergence easier. + if(beam_misconvergence) + { + const float3 convergence_offsets_rgb = + get_convergence_offsets_x_vector(); + const float3 offset_u_rgb = + convergence_offsets_rgb * texture_size_inv.xxx; + const float2 scanline_uv_r = tex_uv - float2(offset_u_rgb.r, 0.0); + const float2 scanline_uv_g = tex_uv - float2(offset_u_rgb.g, 0.0); + const float2 scanline_uv_b = tex_uv - float2(offset_u_rgb.b, 0.0); + const float3 sample_r = sample_single_scanline_horizontal( + tex, scanline_uv_r, tex_size, texture_size_inv); + const float3 sample_g = sample_single_scanline_horizontal( + tex, scanline_uv_g, tex_size, texture_size_inv); + const float3 sample_b = sample_single_scanline_horizontal( + tex, scanline_uv_b, tex_size, texture_size_inv); + return float3(sample_r.r, sample_g.g, sample_b.b); + } + else + { + return sample_single_scanline_horizontal(tex, tex_uv, tex_size, + texture_size_inv); + } +} + +float2 get_last_scanline_uv(const float2 tex_uv, const float2 tex_size, + const float2 texture_size_inv, const float2 il_step_multiple, + const float frame_count, out float dist) +{ + // Compute texture coords for the last/upper scanline, accounting for + // interlacing: With interlacing, only consider even/odd scanlines every + // other frame. Top-field first (TFF) order puts even scanlines on even + // frames, and BFF order puts them on odd frames. Texels are centered at: + // frac(tex_uv * tex_size) == x.5 + // Caution: If these coordinates ever seem incorrect, first make sure it's + // not because anisotropic filtering is blurring across field boundaries. + // Note: TFF/BFF won't matter for sources that double-weave or similar. + // wtf fixme +// const float interlace_bff1 = 1.0; + const float field_offset = floor(il_step_multiple.y * 0.75) * + fmod(frame_count + float(interlace_bff), 2.0); + const float2 curr_texel = tex_uv * tex_size; + // Use under_half to fix a rounding bug right around exact texel locations. + const float2 prev_texel_num = floor(curr_texel - float2(under_half)); + const float wrong_field = fmod( + prev_texel_num.y + field_offset, il_step_multiple.y); + const float2 scanline_texel_num = prev_texel_num - float2(0.0, wrong_field); + // Snap to the center of the previous scanline in the current field: + const float2 scanline_texel = scanline_texel_num + float2(0.5); + const float2 scanline_uv = scanline_texel * texture_size_inv; + // Save the sample's distance from the scanline, in units of scanlines: + dist = (curr_texel.y - scanline_texel.y)/il_step_multiple.y; + return scanline_uv; +} + +inline bool is_interlaced(float num_lines) +{ + // Detect interlacing based on the number of lines in the source. + if(interlace_detect) + { + // NTSC: 525 lines, 262.5/field; 486 active (2 half-lines), 243/field + // NTSC Emulators: Typically 224 or 240 lines + // PAL: 625 lines, 312.5/field; 576 active (typical), 288/field + // PAL Emulators: ? + // ATSC: 720p, 1080i, 1080p + // Where do we place our cutoffs? Assumptions: + // 1.) We only need to care about active lines. + // 2.) Anything > 288 and <= 576 lines is probably interlaced. + // 3.) Anything > 576 lines is probably not interlaced... + // 4.) ...except 1080 lines, which is a crapshoot (user decision). + // 5.) Just in case the main program uses calculated video sizes, + // we should nudge the float thresholds a bit. + const bool sd_interlace = ((num_lines > 288.5) && (num_lines < 576.5)); + const bool hd_interlace = bool(interlace_1080i) ? + ((num_lines > 1079.5) && (num_lines < 1080.5)) : + false; + return (sd_interlace || hd_interlace); + } + else + { + return false; + } +} + +#endif // SCANLINE_FUNCTIONS_H + +///////////////////////////// END SCANLINE-FUNCTIONS //////////////////////////// + +//#include "../../../../include/gamma-management.h" + +//////////////////////////// BEGIN GAMMA-MANAGEMENT ////////////////////////// + +#ifndef GAMMA_MANAGEMENT_H +#define GAMMA_MANAGEMENT_H + +///////////////////////////////// MIT LICENSE //////////////////////////////// + +// Copyright (C) 2014 TroggleMonkey +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to +// deal in the Software without restriction, including without limitation the +// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +// sell copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +// IN THE SOFTWARE. + +///////////////////////////////// DESCRIPTION //////////////////////////////// + +// This file provides gamma-aware tex*D*() and encode_output() functions. +// Requires: Before #include-ing this file, the including file must #define +// the following macros when applicable and follow their rules: +// 1.) #define FIRST_PASS if this is the first pass. +// 2.) #define LAST_PASS if this is the last pass. +// 3.) If sRGB is available, set srgb_framebufferN = "true" for +// every pass except the last in your .cgp preset. +// 4.) If sRGB isn't available but you want gamma-correctness with +// no banding, #define GAMMA_ENCODE_EVERY_FBO each pass. +// 5.) #define SIMULATE_CRT_ON_LCD if desired (precedence over 5-7) +// 6.) #define SIMULATE_GBA_ON_LCD if desired (precedence over 6-7) +// 7.) #define SIMULATE_LCD_ON_CRT if desired (precedence over 7) +// 8.) #define SIMULATE_GBA_ON_CRT if desired (precedence over -) +// If an option in [5, 8] is #defined in the first or last pass, it +// should be #defined for both. It shouldn't make a difference +// whether it's #defined for intermediate passes or not. +// Optional: The including file (or an earlier included file) may optionally +// #define a number of macros indicating it will override certain +// macros and associated constants are as follows: +// static constants with either static or uniform constants. The +// 1.) OVERRIDE_STANDARD_GAMMA: The user must first define: +// static const float ntsc_gamma +// static const float pal_gamma +// static const float crt_reference_gamma_high +// static const float crt_reference_gamma_low +// static const float lcd_reference_gamma +// static const float crt_office_gamma +// static const float lcd_office_gamma +// 2.) OVERRIDE_DEVICE_GAMMA: The user must first define: +// static const float crt_gamma +// static const float gba_gamma +// static const float lcd_gamma +// 3.) OVERRIDE_FINAL_GAMMA: The user must first define: +// static const float input_gamma +// static const float intermediate_gamma +// static const float output_gamma +// (intermediate_gamma is for GAMMA_ENCODE_EVERY_FBO.) +// 4.) OVERRIDE_ALPHA_ASSUMPTIONS: The user must first define: +// static const bool assume_opaque_alpha +// The gamma constant overrides must be used in every pass or none, +// and OVERRIDE_FINAL_GAMMA bypasses all of the SIMULATE* macros. +// OVERRIDE_ALPHA_ASSUMPTIONS may be set on a per-pass basis. +// Usage: After setting macros appropriately, ignore gamma correction and +// replace all tex*D*() calls with equivalent gamma-aware +// tex*D*_linearize calls, except: +// 1.) When you read an LUT, use regular tex*D or a gamma-specified +// function, depending on its gamma encoding: +// tex*D*_linearize_gamma (takes a runtime gamma parameter) +// 2.) If you must read pass0's original input in a later pass, use +// tex2D_linearize_ntsc_gamma. If you want to read pass0's +// input with gamma-corrected bilinear filtering, consider +// creating a first linearizing pass and reading from the input +// of pass1 later. +// Then, return encode_output(color) from every fragment shader. +// Finally, use the global gamma_aware_bilinear boolean if you want +// to statically branch based on whether bilinear filtering is +// gamma-correct or not (e.g. for placing Gaussian blur samples). +// +// Detailed Policy: +// tex*D*_linearize() functions enforce a consistent gamma-management policy +// based on the FIRST_PASS and GAMMA_ENCODE_EVERY_FBO settings. They assume +// their input texture has the same encoding characteristics as the input for +// the current pass (which doesn't apply to the exceptions listed above). +// Similarly, encode_output() enforces a policy based on the LAST_PASS and +// GAMMA_ENCODE_EVERY_FBO settings. Together, they result in one of the +// following two pipelines. +// Typical pipeline with intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = linear_color; // Automatic sRGB encoding +// linear_color = intermediate_output; // Automatic sRGB decoding +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Typical pipeline without intermediate sRGB framebuffers: +// linear_color = pow(pass0_encoded_color, input_gamma); +// intermediate_output = pow(linear_color, 1.0/intermediate_gamma); +// linear_color = pow(intermediate_output, intermediate_gamma); +// final_output = pow(intermediate_output, 1.0/output_gamma); +// Using GAMMA_ENCODE_EVERY_FBO is much slower, but it's provided as a way to +// easily get gamma-correctness without banding on devices where sRGB isn't +// supported. +// +// Use This Header to Maximize Code Reuse: +// The purpose of this header is to provide a consistent interface for texture +// reads and output gamma-encoding that localizes and abstracts away all the +// annoying details. This greatly reduces the amount of code in each shader +// pass that depends on the pass number in the .cgp preset or whether sRGB +// FBO's are being used: You can trivially change the gamma behavior of your +// whole pass by commenting or uncommenting 1-3 #defines. To reuse the same +// code in your first, Nth, and last passes, you can even put it all in another +// header file and #include it from skeleton .cg files that #define the +// appropriate pass-specific settings. +// +// Rationale for Using Three Macros: +// This file uses GAMMA_ENCODE_EVERY_FBO instead of an opposite macro like +// SRGB_PIPELINE to ensure sRGB is assumed by default, which hopefully imposes +// a lower maintenance burden on each pass. At first glance it seems we could +// accomplish everything with two macros: GAMMA_CORRECT_IN / GAMMA_CORRECT_OUT. +// This works for simple use cases where input_gamma == output_gamma, but it +// breaks down for more complex scenarios like CRT simulation, where the pass +// number determines the gamma encoding of the input and output. + + +/////////////////////////////// BASE CONSTANTS /////////////////////////////// + +// Set standard gamma constants, but allow users to override them: +#ifndef OVERRIDE_STANDARD_GAMMA + // Standard encoding gammas: + static const float ntsc_gamma = 2.2; // Best to use NTSC for PAL too? + static const float pal_gamma = 2.8; // Never actually 2.8 in practice + // Typical device decoding gammas (only use for emulating devices): + // CRT/LCD reference gammas are higher than NTSC and Rec.709 video standard + // gammas: The standards purposely undercorrected for an analog CRT's + // assumed 2.5 reference display gamma to maintain contrast in assumed + // [dark] viewing conditions: http://www.poynton.com/PDFs/GammaFAQ.pdf + // These unstated assumptions about display gamma and perceptual rendering + // intent caused a lot of confusion, and more modern CRT's seemed to target + // NTSC 2.2 gamma with circuitry. LCD displays seem to have followed suit + // (they struggle near black with 2.5 gamma anyway), especially PC/laptop + // displays designed to view sRGB in bright environments. (Standards are + // also in flux again with BT.1886, but it's underspecified for displays.) + static const float crt_reference_gamma_high = 2.5; // In (2.35, 2.55) + static const float crt_reference_gamma_low = 2.35; // In (2.35, 2.55) + static const float lcd_reference_gamma = 2.5; // To match CRT + static const float crt_office_gamma = 2.2; // Circuitry-adjusted for NTSC + static const float lcd_office_gamma = 2.2; // Approximates sRGB +#endif // OVERRIDE_STANDARD_GAMMA + +// Assuming alpha == 1.0 might make it easier for users to avoid some bugs, +// but only if they're aware of it. +#ifndef OVERRIDE_ALPHA_ASSUMPTIONS + static const bool assume_opaque_alpha = false; +#endif + + +/////////////////////// DERIVED CONSTANTS AS FUNCTIONS /////////////////////// + +// gamma-management.h should be compatible with overriding gamma values with +// runtime user parameters, but we can only define other global constants in +// terms of static constants, not uniform user parameters. To get around this +// limitation, we need to define derived constants using functions. + +// Set device gamma constants, but allow users to override them: +#ifdef OVERRIDE_DEVICE_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_crt_gamma() { return crt_gamma; } + inline float get_gba_gamma() { return gba_gamma; } + inline float get_lcd_gamma() { return lcd_gamma; } +#else + inline float get_crt_gamma() { return crt_reference_gamma_high; } + inline float get_gba_gamma() { return 3.5; } // Game Boy Advance; in (3.0, 4.0) + inline float get_lcd_gamma() { return lcd_office_gamma; } +#endif // OVERRIDE_DEVICE_GAMMA + +// Set decoding/encoding gammas for the first/lass passes, but allow overrides: +#ifdef OVERRIDE_FINAL_GAMMA + // The user promises to globally define the appropriate constants: + inline float get_intermediate_gamma() { return intermediate_gamma; } + inline float get_input_gamma() { return input_gamma; } + inline float get_output_gamma() { return output_gamma; } +#else + // If we gamma-correct every pass, always use ntsc_gamma between passes to + // ensure middle passes don't need to care if anything is being simulated: + inline float get_intermediate_gamma() { return ntsc_gamma; } + #ifdef SIMULATE_CRT_ON_LCD + inline float get_input_gamma() { return get_crt_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_LCD + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_lcd_gamma(); } + #else + #ifdef SIMULATE_LCD_ON_CRT + inline float get_input_gamma() { return get_lcd_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else + #ifdef SIMULATE_GBA_ON_CRT + inline float get_input_gamma() { return get_gba_gamma(); } + inline float get_output_gamma() { return get_crt_gamma(); } + #else // Don't simulate anything: + inline float get_input_gamma() { return ntsc_gamma; } + inline float get_output_gamma() { return ntsc_gamma; } + #endif // SIMULATE_GBA_ON_CRT + #endif // SIMULATE_LCD_ON_CRT + #endif // SIMULATE_GBA_ON_LCD + #endif // SIMULATE_CRT_ON_LCD +#endif // OVERRIDE_FINAL_GAMMA + +// Set decoding/encoding gammas for the current pass. Use static constants for +// linearize_input and gamma_encode_output, because they aren't derived, and +// they let the compiler do dead-code elimination. +#ifndef GAMMA_ENCODE_EVERY_FBO + #ifdef FIRST_PASS + static const bool linearize_input = true; + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + static const bool linearize_input = false; + inline float get_pass_input_gamma() { return 1.0; } + #endif + #ifdef LAST_PASS + static const bool gamma_encode_output = true; + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + static const bool gamma_encode_output = false; + inline float get_pass_output_gamma() { return 1.0; } + #endif +#else + static const bool linearize_input = true; + static const bool gamma_encode_output = true; + #ifdef FIRST_PASS + inline float get_pass_input_gamma() { return get_input_gamma(); } + #else + inline float get_pass_input_gamma() { return get_intermediate_gamma(); } + #endif + #ifdef LAST_PASS + inline float get_pass_output_gamma() { return get_output_gamma(); } + #else + inline float get_pass_output_gamma() { return get_intermediate_gamma(); } + #endif +#endif + +// Users might want to know if bilinear filtering will be gamma-correct: +static const bool gamma_aware_bilinear = !linearize_input; + + +////////////////////// COLOR ENCODING/DECODING FUNCTIONS ///////////////////// + +inline float4 encode_output(const float4 color) +{ + if(gamma_encode_output) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(1.0/get_pass_output_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_input(const float4 color) +{ + if(linearize_input) + { + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), 1.0); + } + else + { + return float4(pow(color.rgb, float3(get_pass_input_gamma())), color.a); + } + } + else + { + return color; + } +} + +inline float4 decode_gamma_input(const float4 color, const float3 gamma) +{ + if(assume_opaque_alpha) + { + return float4(pow(color.rgb, gamma), 1.0); + } + else + { + return float4(pow(color.rgb, gamma), color.a); + } +} + +//TODO/FIXME: I have no idea why replacing the lookup wrappers with this macro fixes the blurs being offset ¯\_(ツ)_/¯ +//#define tex2D_linearize(C, D) decode_input(vec4(COMPAT_TEXTURE(C, D))) +// EDIT: it's the 'const' in front of the coords that's doing it + +/////////////////////////// TEXTURE LOOKUP WRAPPERS ////////////////////////// + +// "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a wide array of linearizing texture lookup wrapper functions. The +// Cg shader spec Retroarch uses only allows for 2D textures, but 1D and 3D +// lookups are provided for completeness in case that changes someday. Nobody +// is likely to use the *fetch and *proj functions, but they're included just +// in case. The only tex*D texture sampling functions omitted are: +// - tex*Dcmpbias +// - tex*Dcmplod +// - tex*DARRAY* +// - tex*DMS* +// - Variants returning integers +// Standard line length restrictions are ignored below for vertical brevity. +/* +// tex1D: +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1D(tex, tex_coords)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy) +{ return decode_input(tex1D(tex, tex_coords, dx, dy)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +inline float4 tex1D_linearize(const sampler1D tex, const float2 tex_coords, const float dx, const float dy, const int texel_off) +{ return decode_input(tex1D(tex, tex_coords, dx, dy, texel_off)); } + +// tex1Dbias: +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dbias(tex, tex_coords)); } + +inline float4 tex1Dbias_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dbias(tex, tex_coords, texel_off)); } + +// tex1Dfetch: +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords) +{ return decode_input(tex1Dfetch(tex, tex_coords)); } + +inline float4 tex1Dfetch_linearize(const sampler1D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex1Dfetch(tex, tex_coords, texel_off)); } + +// tex1Dlod: +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords) +{ return decode_input(tex1Dlod(tex, tex_coords)); } + +inline float4 tex1Dlod_linearize(const sampler1D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex1Dlod(tex, tex_coords, texel_off)); } + +// tex1Dproj: +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords) +{ return decode_input(tex1Dproj(tex, tex_coords)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float2 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex1Dproj_linearize(const sampler1D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex1Dproj(tex, tex_coords, texel_off)); } +*/ +// tex2D: +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords) +{ return decode_input(COMPAT_TEXTURE(tex, tex_coords.xy)); } + +inline float4 tex2D_linearize(const sampler2D tex, float2 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords, texel_off)); } + +inline float4 tex2D_linearize(const sampler2D tex, float3 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +//inline float4 tex2D_linearize(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off) +//{ return decode_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off)); } + +// tex2Dbias: +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords) +//{ return decode_input(tex2Dbias(tex, tex_coords)); } + +//inline float4 tex2Dbias_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dbias(tex, tex_coords, texel_off)); } + +// tex2Dfetch: +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords) +//{ return decode_input(tex2Dfetch(tex, tex_coords)); } + +//inline float4 tex2Dfetch_linearize(const sampler2D tex, const int4 tex_coords, const int texel_off) +//{ return decode_input(tex2Dfetch(tex, tex_coords, texel_off)); } + +// tex2Dlod: +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords) +{ return decode_input(textureLod(tex, tex_coords.xy, 0.0)); } + +inline float4 tex2Dlod_linearize(const sampler2D tex, float4 tex_coords, int texel_off) +{ return decode_input(textureLod(tex, tex_coords.xy, texel_off)); } +/* +// tex2Dproj: +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords) +{ return decode_input(tex2Dproj(tex, tex_coords)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } + +inline float4 tex2Dproj_linearize(const sampler2D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex2Dproj(tex, tex_coords, texel_off)); } +*/ +/* +// tex3D: +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords) +{ return decode_input(tex3D(tex, tex_coords)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, texel_off)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy) +{ return decode_input(tex3D(tex, tex_coords, dx, dy)); } + +inline float4 tex3D_linearize(const sampler3D tex, const float3 tex_coords, const float3 dx, const float3 dy, const int texel_off) +{ return decode_input(tex3D(tex, tex_coords, dx, dy, texel_off)); } + +// tex3Dbias: +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dbias(tex, tex_coords)); } + +inline float4 tex3Dbias_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dbias(tex, tex_coords, texel_off)); } + +// tex3Dfetch: +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords) +{ return decode_input(tex3Dfetch(tex, tex_coords)); } + +inline float4 tex3Dfetch_linearize(const sampler3D tex, const int4 tex_coords, const int texel_off) +{ return decode_input(tex3Dfetch(tex, tex_coords, texel_off)); } + +// tex3Dlod: +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dlod(tex, tex_coords)); } + +inline float4 tex3Dlod_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dlod(tex, tex_coords, texel_off)); } + +// tex3Dproj: +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords) +{ return decode_input(tex3Dproj(tex, tex_coords)); } + +inline float4 tex3Dproj_linearize(const sampler3D tex, const float4 tex_coords, const int texel_off) +{ return decode_input(tex3Dproj(tex, tex_coords, texel_off)); } +/////////* + +// NONSTANDARD "SMART" LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// This narrow selection of nonstandard tex2D* functions can be useful: + +// tex2Dlod0: Automatically fill in the tex2D LOD parameter for mip level 0. +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0))); } + +//inline float4 tex2Dlod0_linearize(const sampler2D tex, const float2 tex_coords, const int texel_off) +//{ return decode_input(tex2Dlod(tex, float4(tex_coords, 0.0, 0.0), texel_off)); } + + +// MANUALLY LINEARIZING TEXTURE LOOKUP FUNCTIONS: +// Provide a narrower selection of tex2D* wrapper functions that decode an +// input sample with a specified gamma value. These are useful for reading +// LUT's and for reading the input of pass0 in a later pass. + +// tex2D: +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords), gamma); } + +inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float3 gamma) +{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords.xy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float2 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } + +//inline float4 tex2D_linearize_gamma(const sampler2D tex, const float3 tex_coords, const float2 dx, const float2 dy, const int texel_off, const float3 gamma) +//{ return decode_gamma_input(COMPAT_TEXTURE(tex, tex_coords, dx, dy, texel_off), gamma); } +/* +// tex2Dbias: +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords), gamma); } + +inline float4 tex2Dbias_linearize_gamma(const sampler2D tex, const float4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dbias(tex, tex_coords, texel_off), gamma); } + +// tex2Dfetch: +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords), gamma); } + +inline float4 tex2Dfetch_linearize_gamma(const sampler2D tex, const int4 tex_coords, const int texel_off, const float3 gamma) +{ return decode_gamma_input(tex2Dfetch(tex, tex_coords, texel_off), gamma); } +*/ +// tex2Dlod: +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, 0.0), gamma); } + +inline float4 tex2Dlod_linearize_gamma(const sampler2D tex, float4 tex_coords, int texel_off, float3 gamma) +{ return decode_gamma_input(textureLod(tex, tex_coords.xy, texel_off), gamma); } + + +#endif // GAMMA_MANAGEMENT_H + +//////////////////////////// END GAMMA-MANAGEMENT ////////////////////////// + +#undef COMPAT_PRECISION +#undef COMPAT_TEXTURE + +void main() { + gl_Position = position; + vTexCoord = texCoord * 1.0001; + + // Detect interlacing: il_step_multiple indicates the step multiple between + // lines: 1 is for progressive sources, and 2 is for interlaced sources. + float2 video_size_ = video_size.xy; + const float y_step = 1.0 + float(is_interlaced(video_size_.y)); + il_step_multiple = float2(1.0, y_step); + // Get the uv tex coords step between one texel (x) and scanline (y): + uv_step = il_step_multiple / texture_size; + + // If shader parameters are used, {min, max}_{sigma, shape} are runtime + // values. Compute {sigma, shape}_range outside of scanline_contrib() so + // they aren't computed once per scanline (6 times per fragment and up to + // 18 times per vertex): + // TODO/FIXME: if these aren't used, why are they calculated? commenting for now +// const floatsigma_range = max(beam_max_sigma, beam_min_sigma) - +// beam_min_sigma; +// const float shape_range = max(beam_max_shape, beam_min_shape) - +// beam_min_shape; + + // We need the pixel height in scanlines for antialiased/integral sampling: + const float ph = (video_size_.y / output_size.y) / + il_step_multiple.y; + pixel_height_in_scanlines = ph; +} \ No newline at end of file diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearApertureGrille15Wide8And5d5Spacing.png b/shaders/CRT-Royale.shader/textures/TileableLinearApertureGrille15Wide8And5d5Spacing.png new file mode 100644 index 000000000..2995ae5f4 Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearApertureGrille15Wide8And5d5Spacing.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearApertureGrille15Wide8And5d5SpacingResizeTo64.png b/shaders/CRT-Royale.shader/textures/TileableLinearApertureGrille15Wide8And5d5SpacingResizeTo64.png new file mode 100644 index 000000000..2c3f21eed Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearApertureGrille15Wide8And5d5SpacingResizeTo64.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearShadowMask.png b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMask.png new file mode 100644 index 000000000..ca4095649 Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMask.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskEDP.png b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskEDP.png new file mode 100644 index 000000000..a3844dc2a Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskEDP.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskEDPResizeTo64.png b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskEDPResizeTo64.png new file mode 100644 index 000000000..b61d92a0e Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskEDPResizeTo64.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskResizeTo64.png b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskResizeTo64.png new file mode 100644 index 000000000..9b66ffba3 Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearShadowMaskResizeTo64.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing.png b/shaders/CRT-Royale.shader/textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing.png new file mode 100644 index 000000000..eb20b2316 Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacing.png differ diff --git a/shaders/CRT-Royale.shader/textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacingResizeTo64.png b/shaders/CRT-Royale.shader/textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacingResizeTo64.png new file mode 100644 index 000000000..df518db57 Binary files /dev/null and b/shaders/CRT-Royale.shader/textures/TileableLinearSlotMaskTall15Wide9And4d5Horizontal9d14VerticalSpacingResizeTo64.png differ