[YuutaTsubasa]

Issue #224: Galaxy A55 (SM-A556E, Exynos 1480, Xclipse 530, driver 24.0.539) stays at "PREPARING PIPELINES 0 OF 455" and Android eventually kills the game for low memory (exit reason 3).

Cause

Firebase Test Lab's a55xzh (Galaxy A55, Android 16, Xclipse 530, driver 24.0.545) reproduces it with the 0.10.14 APK's graphics self test. The driver's shader compiler (logcat tag GSC) spends 209 s on the first recipe's vertex shader (eaf61b182ad92401, which has XenosRecomp's while(true) switch(pc) dispatch loop). Resident memory grows past 550 MB, then the compiler asserts in SCAssembler.cpp SCGenerateLongBranch and Compiler.cpp Compile. On the player's phone the low-memory killer ends the game before that.

Bisection on the A55 (self-test scenarios 9-13, 60 s limit per recipe):

Scenario	Result
as the device says	stuck at recipe 0, 550 MB
SFR_TEST_DISPATCH_LOOP=1	455/455 in 14.2 s, slowest recipe 253 ms
SFR_TEST_NO_DESCRIPTOR_INDEXING=1	stuck at recipe 0
SFR_TEST_NO_DEVICE_ADDRESS=1	stuck at recipe 0
SFR_TEST_NO_CLIP_DISTANCE=1	stuck at recipe 0

Fix

• The dispatch-loop rewrite (spirv_dispatch_loop.h, already on for Apple since Take XenosRecomp's dispatch loops out of the shaders on Apple (M1 race 6.5 to 60 fps) #192) is now also on for Vulkan vendor 0x144d (Samsung Xclipse). SFR_TEST_DISPATCH_LOOP=0 still keeps the loops (launcher Test options: "Rewrite shader dispatch loops: Off"). Other GPUs are unchanged.
• Pipeline preparation can now be diagnosed from game.log:
  • A recipe that takes more than 5 s is logged as PIPELINE_PREWARM slow recipe=<n> of=<total> step=shaders|pipeline vs= ps= seconds= resident_mb=, and logged again at 10, 20, 40 s and so on.
  • A progress= line is written every 10 s.
  • The complete= line now names the slowest recipe.
  • On Android these lines also go to logcat.
  • SFR_PIPELINE_PREWARM_TRACE=1 logs every recipe. SFR_PIPELINE_PREWARM_TIMEOUT=<s> ends the process when one recipe takes longer than that; it is for the self test.
• Self-test scenarios 9-14 (GraphicsTestActivity) only prepare the pipelines, with tracing and the timeout on. Each scenario tries one shader rewrite.
• The test-option hint now mentions Xclipse. docs/vulkan.md has the details.

Verification

Firebase Test Lab, APK from this branch, scenarios 9 (as the device says), 3 (draw with every pipeline) and 14 (loops forced back on):

Device	Scenario 9	Scenario 3	Scenario 14
Galaxy A55 (a55xzh, Xclipse 530)	455/455 in 18.2 s	455 drawn, no hang, ok=1	stuck at recipe 0 again (as expected)
Galaxy S24 (SC-51E, Adreno 750)	455/455 in 9.2 s	455 drawn, ok=1	455/455
Pixel 10a (stallion, Mali-G715)	455/455 in 10.3 s	455 drawn, ok=1	455/455

Local checks:

• Windows: ctest 163/165 passed. native_shader_abi and native_resolution are not built by build_tools, and CI excludes them too.
• Linux (WSL, lavapipe): ctest 159/159 passed.
• Python unittest: 356 tests OK, 11 skipped.
• The Android arm64 build passes.

Not tested

• A real race on an Xclipse device: picture and frame rate with the rewrite on a non-Apple GPU. The self test's draws only show that every pipeline draws without a hang.
• Other Xclipse GPUs (920/940/540) and other driver versions.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.