GBA: Optimise framebuffer clearing

Measuring in mgBA, time taken reduced from ~52 ms to ~3ms

misc/gba/Makefile

@@ -36,15 +36,11 @@ DEPFILES := $(OBJS:%.o=%.d)
 # Code generation
 #---------------------------------------------------------------------------------
 ARCH	= -mthumb -mthumb-interwork
-CFLAGS	= -g -Wall -O2 -DPLAT_GBA -ffunction-sections -fdata-sections -mcpu=arm7tdmi -mtune=arm7tdmi $(ARCH)
+CFLAGS	= -g -gdwarf-4 -Wall -O2 -DPLAT_GBA -ffunction-sections -fdata-sections -mcpu=arm7tdmi -mtune=arm7tdmi $(ARCH)
 ASFLAGS	= -g $(ARCH)
 
 LDFLAGS	= -specs=gba.specs -g $(ARCH)
 
-LIBGBA		:= $(DEVKITPRO)/libgba
-INCLUDES 	+= $(foreach dir, $(LIBGBA), -I$(dir)/include)
-LDFLAGS 	+= $(foreach dir, $(LIBGBA), -L$(dir)/lib)
-
 
 #---------------------------------------------------------------------------------
 # Compiler tools

src/Graphics_SoftMin.c

@@ -129,6 +129,11 @@ void Gfx_SetAlphaArgBlend(cc_bool enabled) { }
 static void ClearColorBuffer(void) {
 	int i, x, y, size = fb_width * fb_height;
 
+#ifdef CC_BUILD_GBA
+	/* in mGBA, fast clear takes ~2ms compared to ~52ms of code below */
+	extern void VRAM_FastClear(BitmapCol color);
+	VRAM_FastClear(clearColor);
+#else
 	if (cb_stride == fb_width) {
 		for (i = 0; i < size; i++) colorBuffer[i] = clearColor;
 	} else {
@@ -140,6 +145,7 @@ static void ClearColorBuffer(void) {
 			}
 		}
 	}
+#endif
 }
 
 void Gfx_ClearBuffers(GfxBuffers buffers) {

src/gba/AsmFuncs.S

@@ -0,0 +1,79 @@
+
+#define BEG_THUMB_FUNC(name_) \
+	.global name_; \
+	.thumb_func; \
+	.align 2; \
+	name_:
+
+#define BEG_ARM_FUNC(name_) \
+	.global name_; \
+	.arm; \
+	.align 4; \
+	name_:
+
+#define END_FUNC(name_) \
+	.size name_, . - name_; \
+	.type name_, %function;
+
+
+// ===============================
+// FUNCTIONS IN TEXT SECTION
+// ===============================
+.section text
+
+
+// ===============================
+// FUNCTIONS IN IWRAM SECTION
+// ===============================
+.section .iwram,"ax",%progbits
+
+#define R_CUR r0 // r0 = beg address (incremented in function)
+#define R_END r1 // r1 = end address
+#define R_VAL r2 // r2 = value to fill
+
+BEG_ARM_FUNC(fastset_256_bytes)
+	// Spill callee saved registers + LR onto stack
+	stmfd sp!, {r4-r9, lr}
+
+	// Clone 'value' for 'store multi CPU registers' loop
+	mov	r3, R_VAL
+	mov	r4, R_VAL
+	mov r5, R_VAL
+	mov r6, R_VAL
+	mov r7, R_VAL
+	mov r8, R_VAL
+	mov r9, R_VAL
+
+fastset_loop:
+	cmp R_CUR, R_END
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	stmltia R_CUR!, {r2-r9}
+	blt fastset_loop
+
+	// Restore saved CPU registers + set PC to LR
+	ldmfd  sp!, {r4-r9, pc}
+END_FUNC(fastset_256_bytes)
+
+
+// ===============================
+// FUNCTIONS IN EWRAM SECTION
+// ===============================
+.section .ewram,"ax",%progbits
+BEG_THUMB_FUNC(nocash_log)
+	// nocash looks for this specific pattern
+	mov 	r12, r12
+	b		nocash_log_return
+	.short	0x6464, 0x0000
+.global nocash_msg // Platform_GBA writes directly to this
+nocash_msg:
+	.space	82
+
+nocash_log_return:
+	bx		lr
+END_FUNC(nocash_log)

src/gba/NocashLog.S

@@ -1,22 +0,0 @@
-.global nocash_log
-.global nocash_msg
-
-// BEG nocash_log
-.section .ewram,"ax",%progbits
-.thumb_func
-.align 2
-
-// nocash looks for this specific pattern
-nocash_log:
-	mov 	r12, r12
-	b		nocash_log_return
-	.short	0x6464, 0x0000
-nocash_msg:
-	.space	82
-
-nocash_log_return:
-	bx		lr
-
-.size nocash_log, .-nocash_log
-.type nocash_log, %function
-// END nocash_log

src/gba/Platform_GBA.c

@@ -13,9 +13,8 @@
 #include "../Options.h"
 #include "../Animations.h"
 
-#include <stdlib.h>
+#include <stdint.h>
 #include <string.h>
-#include <unistd.h>
 #include "gbadefs.h"
 
 #include "../../third_party/tinyalloc/tinyalloc.c"
@@ -203,9 +202,10 @@ void DateTime_CurrentLocal(struct cc_datetime* t) {
 void CrashHandler_Install(void) {
 }
 
+
 void Process_Abort2(cc_result result, const char* raw_msg) {
 	Platform_LogConst(raw_msg);
-	_exit(0);
+	Process_Exit(0);
 }
 
 
@@ -333,5 +333,10 @@ cc_result Platform_SetDefaultCurrentDirectory(int argc, char **argv) {
 	return 0; 
 }
 
-void Process_Exit(cc_result code) { _exit(code); }
+extern void bios_soft_reset(void);
+void Process_Exit(cc_result code) { 
+	//*(vu8*)0x03007FFA = 0x00; // controls reset address
+	// TODO jump to start_vector instead?
+	for (;;) { __asm__ volatile(""); }
+}
 

src/gba/Window_GBA.c

@@ -108,17 +108,20 @@ void Gamepads_Init(void) {
 	Input.Sources |= INPUT_SOURCE_GAMEPAD;
 }
 
+#define PAUSE_MASK (KEY_A|KEY_B|KEY_START|KEY_SELECT)
 void Gamepads_Process(float delta) {
 	int port = Gamepad_Connect(0x5BA, pad_defaults);
 	int mods = ~REG_KEYINPUT;
 
+	// TODO see comment in Platform_GBA.c, doesn't work anyways
+	//if ((mods & PAUSE_MASK) == PAUSE_MASK)
+	//	Process_Exit(0);
+	
 	Gamepad_SetButton(port, CCPAD_L, mods & KEY_L);
 	Gamepad_SetButton(port, CCPAD_R, mods & KEY_R);
 	
 	Gamepad_SetButton(port, CCPAD_1, mods & KEY_A);
 	Gamepad_SetButton(port, CCPAD_2, mods & KEY_B);
-	//Gamepad_SetButton(port, CCPAD_3, mods & KEY_X);
-	//Gamepad_SetButton(port, CCPAD_4, mods & KEY_Y);
 	
 	Gamepad_SetButton(port, CCPAD_START,  mods & KEY_START);
 	Gamepad_SetButton(port, CCPAD_SELECT, mods & KEY_SELECT);
@@ -133,6 +136,15 @@ void Gamepads_Process(float delta) {
 /*########################################################################################################################*
 *------------------------------------------------------Framebuffer--------------------------------------------------------*
 *#########################################################################################################################*/
+extern void fastset_256_bytes(char* beg, char* end, int value);
+
+void VRAM_FastClear(BitmapCol col) {
+	int value = (col << 16) | col;
+
+	char* vram = (char*)MEM_VRAM;
+	fastset_256_bytes(vram, vram + SCREEN_WIDTH * SCREEN_HEIGHT * 2, value);
+}
+
 void Window_AllocFramebuffer(struct Bitmap* bmp, int width, int height) {
 	bmp->scan0  = (BitmapCol*)MEM_VRAM;
 	bmp->width  = width;