psxrecomp/docs/SDL_BACKENDS.md

SDL backends

PSXRecomp uses SDL3 by default. A compatible system SDL3 package is preferred; otherwise CMake downloads the integrity-pinned SDL 3.4.10 release. To build the compatibility backend explicitly:

cmake -S . -B build-sdl2 -DPSX_SDL_BACKEND=SDL2

PSX_SDL_BACKEND accepts only SDL3 or SDL2; it never silently changes the requested backend. PSX_SDL3_FETCH=OFF makes an SDL3 build system-package-only.

Initial Windows A/B exercise

These measurements are smoke-level comparisons, not claims that one SDL generation makes emulation faster. Both builds used MinGW GCC 15.2, Release, static SDL, OpenBIOS, the OpenGL renderer, and a 165 Hz Windows host. Each row is three warmed runs of the runtime's frame-window diagnostic:

Game and workload	SDL3 wall time	SDL2 wall time	SDL3 difference
Tomba 2, frames 60–660	9506.5 ± 10.9 ms	9511.3 ± 19.0 ms	-0.05%
Mega Man X6, frames 60–660	8174.0 ± 114.0 ms	8199.2 ± 287.0 ms	-0.31%

Every Tomba 2 run executed exactly 928,243 dirty instructions and 309,407 dirty dispatches. Every Mega Man X6 run executed 3,962,702 dirty instructions and 1,048,509 dirty dispatches. The paced wall-time differences are smaller than the observed variance, so the useful result is parity rather than a speed win.

The SDL3 build was additionally exercised through:

• the shared recomp-ui launcher event loop;
• the SDL renderer presentation path (Tomba 2, 240 frames); and
• the Vulkan renderer on an RTX 3080 Ti, including SDL3 instance-extension and surface creation APIs (Tomba 2, 240 frames).

The repeatable diagnostic setup is:

PSX_RUNTIME_PERF_DIAG=1 \
PSX_BENCH_WINDOW=60:660 \
PSX_DEV_INPUT=0 \
./game-runtime --no-launcher --disc /path/to/game.cue --renderer opengl

Compare only runs with identical frame windows and dirty instruction/dispatch counts. Warm each executable and disc image before collecting results; cold filesystem and overlay-capture state caused much larger swings than the SDL backend in the initial exercise.