fix(shader): share wave64 VGPRs past the private ones between the two halves by wachi35676 · Pull Request #7410 · boykopovar/AnyPS5

wachi35676 · 2026-10-18T05:05:39Z

What

On GFX10, a wave64 program can declare wave-shared VGPRs (RDNA2 ISA 3.6.5): COMPUTE_PGM_RSRC3.SHARED_VGPR_CNT blocks of eight VGPRs placed after the (COMPUTE_PGM_RSRC1.VGPRS + 1) * 4 private ones. Lane N and lane N + 32 share one copy of each.

ACO uses them to swap the halves of a wave64 (emit_permlane64_shared_vgpr) and to emulate a wave-wide ds_bpermute_b32 (emit_bpermute_shared_vgpr).
The flash-attention reductions of llama.cpp's Vulkan backend end with that swap.
The recompiler kept these VGPRs private, so each half read its own stale copy and lost the other half's partial sums.
In PS5LM Plus (PPSA99581) this gave garbage with -fa on -ctk q8_0 -ctv q8_0, and slightly wrong text in other configurations.

Changes:

The driver decodes RSRC1.VGPRS and RSRC3.SHARED_VGPR_CNT for compute dispatches into ShaderComputeStageInfo. Both are part of the cache key, the dispatch key (0x228) and request serialization (version 50).
A VALU write to a shared VGPR also updates lane N ^ 32's copy, through a new HalfSwapU32 IR op.
- When both halves write, the high half wins, because a wave64 VALU instruction issues its low half first (ISA 4.2.1 Table 9, 6.2.4).
- In SPIR-V the partner is the other half in the same invocation, a subgroup shuffle (64-lane subgroups), or the workgroup exchange (spanned waves).
A non-VALU write to a shared VGPR, such as ACO's ds_bpermute_b32, is modeled while an immediate s_mov_b32, s_mov_b64 or s_bfm_b64 earlier in the block has left one half of EXEC empty, as ACO does. With both halves possibly active the ISA gives no order, so it throws.
These also throw: a VALU instruction that reads the shared VGPR it writes, v_writelane_b32, relative VGPR writes in such a program, wave32 programs with shared VGPRs, and host subgroups narrower than 32 lanes.
RADV also sets SHARED_VGPR_CNT in SPI_SHADER_PGM_RSRC2 for graphics stages. Graphics stages are not modeled, so a draw whose program sets it throws. No graphics program set it in Dusklight, flycast, Porpoise, PSSwanStation, PS5LM, Prospero Win, ProsperoAI or ProsperoLight.
The remaining gaps are recorded in docs/dev/debt/shader.md.

Sources:

AMD RDNA2 ISA 3.6.5, 4.2.1 (Table 9) and 6.2.4, and the GPR_ALLOC table (VGPR_SHARED_SIZE in units of 8).
Mesa be738ab0ea44cbddee6f0dcbd3cd9d16d98118ed:
- src/amd/compiler/aco_lower_to_hw_instr.cpp: emit_permlane64_shared_vgpr, emit_bpermute_shared_vgpr and copy_constant_sgpr.
- src/amd/vulkan/radv_shader.c: S_00B848_VGPRS((num_vgprs - 1) / 4) for wave64, S_00B8A0_SHARED_VGPR_CNT(num_shared_vgprs / 8), and the RSRC2 SHARED_VGPR_CNT of the graphics stages.
- src/amd/registers/gfx10.json: field positions.
Not measured on hardware.

Tested

Windows 11, WinLibs GCC 15.2, full build:

ctest -R "agc_|shader|recompiler|guest_|request_serializer" -E video_out (GPU tests on): 1163/1163 passed.
ctest -R "^(libkernel|kernel|nid|patched_libraries|file_streams|libc)" -E "agc_|video_out": 12/12 passed.

After the rebase onto main 3a40945 (serializer version 50): clean full Windows build. request_serializer, agc_shader_memory, agc_driver_wave_shared_vgprs and the wider serializer, dispatch, subgroup, masked-select, stencil and shader-cache set pass 40/40 on NVIDIA and on AMD. ctest -R ^agc_ on NVIDIA: 1122/1122.

Linux (WSL Ubuntu 24.04, GCC 13.3): libSceAgcDriver plus the recompiler, serializer and AGC driver tests pass. agc_driver_wave_shared_vgprs skips on llvmpipe (subgroup size 8).

New tests:

agc_driver_wave_shared_vgprs runs ACO's half swap, an exec-masked low-half write and ACO's wave64 bpermute sequence word for word, and checks the rejections.
- On main's sources it fails with lane 0 read 0x00000000, expected lane 32's 0x00201020, or with lane 6 read 0x00000000, expected lane 33's 0x00211021 for the bpermute part only.
- With this change it passes on both GPUs, also with spanned waves.
agc_driver_prepared_unregistered_graphics checks that a graphics stage with SHARED_VGPR_CNT is rejected.
agc_driver_graphics covers decoding RSRC1/RSRC3 and serialization.
request_serializer replays a version 49 compute request, which has no wave-shared VGPR words. If the reader reads those words from version 37 as before, it fails with RequestSerializer: truncated data.

PS5LM Plus 0.8.4 (PPSA99581) with Qwen2.5-0.5B-Instruct q4_k_m (Apache-2.0), prompt "The capital of France is", 16 tokens, temperature 0. The title is fixed on NVIDIA (RTX 5070 Ti):

main, -fa on -ctk q8_0 -ctv q8_0: orntymology. In the 19th-arthury.10 (3/3).
This PR: Paris. It is the largest city in Europe and the second largest in the world (3/3), the CPU backend's exact answer.
-fa on with f16 K/V, and -fa off, now give the same answer. Before, they gave Paris. It is the largest city in the world by population… and Paris. It is located in the south of France….

On the AMD Radeon iGPU the title is still garbage in every configuration, -fa off included, both before and after this change; that is a separate bug. Only the replay of the captured flash-attention dispatch is correct there: its max relative error against a numpy reference went from 0.99 to 0.031.

An independent review checked:

the EXEC-half tracker, including cross-block state, every EXEC writer and the inline-constant sign extension;
the graphics register fields;
the ACO sequences, word for word;
negative controls for each part.

A regression run of Dusklight, flycast, Porpoise, PSSwanStation, PS5LM, Prospero Win, ProsperoAI and ProsperoLight (60 s each) showed no new errors.

Checklist

🤖 Generated with Claude Code

github-actions · 2026-10-18T05:33:17Z

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 1395	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 1896	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp`
PR 4040	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 4287	`core/shader/recompiler/Translation/src/TranslationContext/OperandAccess.cpp`
PR 5021	`core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `docs/dev/debt/shader.md`
PR 5371	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 5609	`core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 5925	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6146	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 6188	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Draw/DrawDecode.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6194	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 6235	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Draw/DrawDecode.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 6280	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `docs/dev/debt/shader.md`
PR 6290	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp`
PR 6412	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 6416	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp`
PR 6481	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 6488	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 6548	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 6751	`core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp`
PR 6804	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 6808	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 6858	`core/shader/recompiler/Recompiler.cpp`
PR 6924	`core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 6970	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 7085	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `docs/dev/debt/shader.md`
PR 7223	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 7256	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 7392	`core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `docs/dev/debt/shader.md`
PR 7395	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 7522	`core/shader/recompiler/Recompiler.cpp`
PR 7642	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Deserialize.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp`
PR 7668	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 7684	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 7689	`core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/StencilReplay.cpp`
PR 8004	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 8212	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 8246	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 8458	`docs/dev/debt/shader.md`
PR 8939	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9020	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 9183	`docs/dev/debt/shader.md`
PR 9335	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 9356	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp`
PR 9387	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 9457	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9489	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9508	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Deserialize.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp`
PR 9604	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9633	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9671	`docs/dev/debt/shader.md`
PR 9675	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9700	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9712	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9748	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 9749	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9790	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9852	`core/shader/recompiler/Recompiler.hpp`
PR 9875	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9889	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp` `core/shader/recompiler/Translation/src/MemoryInstructions/MemoryResourceHelpers.cpp` `docs/dev/debt/shader.md`
PR 9895	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9897	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 9915	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 9969	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 10000	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 10049	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 10103	`core/libs/prx/libSceAgcDriver/tests/execution/PreparedUnregisteredGraphics.cpp`
PR 10141	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 10193	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Deserialize.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp`
PR 10233	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 10249	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderInputState.cpp`
PR 10351	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/StencilReplay.cpp`
PR 10359	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/Translation/include/Translation/InstructionTranslator.hpp` `core/shader/recompiler/Translation/src/InstructionTranslator.cpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp`
PR 10472	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 10489	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 10556	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 10652	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 10757	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 10779	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 10867	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvAnalysis.cpp` `docs/dev/debt/shader.md`
PR 11265	`core/shader/recompiler/IntermediateRepresentation/include/IntermediateRepresentation/IrOpcode.hpp` `core/shader/recompiler/IntermediateRepresentation/src/IrOpcode.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterInstructions.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `core/shader/recompiler/Translation/include/Translation/InstructionTranslator.hpp` `core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp` `core/shader/recompiler/Translation/src/InstructionTranslator.cpp`
PR 11351	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp`
PR 11483	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 11490	`core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp` `core/shader/recompiler/Translation/src/TranslationContext/OperandAccess.cpp`
PR 11551	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Draw/DrawDecode.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp`
PR 11557	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `docs/dev/debt/shader.md`
PR 11566	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 11825	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 11998	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 12002	`core/shader/recompiler/IntermediateRepresentation/src/IrOpcode.cpp`
PR 12066	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 12125	`core/shader/recompiler/Translation/src/InstructionTranslator.cpp`
PR 12132	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 12316	`docs/dev/debt/shader.md`
PR 12324	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp`
PR 12458	`core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp` `core/shader/recompiler/Translation/src/ControlFlowInstructions.cpp` `docs/dev/debt/shader.md`
PR 12548	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/IntermediateRepresentation/include/IntermediateRepresentation/IrOpcode.hpp` `core/shader/recompiler/IntermediateRepresentation/src/IrOpcode.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterInstructions.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `core/shader/recompiler/Translation/src/TranslationContext/OperandAccess.cpp` `docs/dev/debt/shader.md`
PR 12662	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 12764	`core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `docs/dev/debt/shader.md`
PR 12790	`core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 12828	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 12840	`core/shader/recompiler/Recompiler.cpp`
PR 12855	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 12944	`core/shader/recompiler/IntermediateRepresentation/src/IrOpcode.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvAnalysis.cpp` `docs/dev/debt/shader.md`
PR 12980	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp` `docs/dev/debt/shader.md`
PR 13276	`core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp`
PR 13855	`core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp`
PR 13863	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp`
PR 13920	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp`
PR 13928	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Deserialize.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp`
PR 13929	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Dispatch/Dispatch.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Deserialize.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp`
PR 13991	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 14032	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 14068	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp` `docs/dev/debt/shader.md`
PR 14072	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvAlu/SpirvAluEmitterMath.cpp` `docs/dev/debt/shader.md`
PR 14076	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 14081	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 14085	`docs/dev/debt/shader.md`
PR 14212	`docs/dev/debt/shader.md`
PR 14217	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 14218	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterInstructions.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `docs/dev/debt/shader.md`
PR 14243	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 14247	`docs/dev/debt/shader.md`
PR 14265	`core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 14266	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.hpp`
PR 14341	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp`
PR 14367	`core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 14377	`docs/dev/debt/shader.md`
PR 14473	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 14561	`core/libs/prx/libSceAgcDriver/Execution/src/Driver/Draw/DrawDecode.cpp` `core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/ControlFlow/src/RequestSerializer.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/tests/RequestSerializerTests.cpp`
PR 14621	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp`
PR 14627	`docs/dev/debt/shader.md`
PR 14628	`docs/dev/debt/shader.md`
PR 14633	`docs/dev/debt/shader.md`
PR 14636	`core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 14692	`docs/dev/debt/shader.md`
PR 14708	`core/shader/recompiler/SpirvBackend/src/SpirvFlowEmitter.cpp` `core/shader/recompiler/Translation/include/Translation/InstructionTranslator.hpp` `core/shader/recompiler/Translation/src/ControlFlowInstructions.cpp` `core/shader/recompiler/Translation/src/DispatchInstructions.cpp` `core/shader/recompiler/Translation/src/InstructionTranslator.cpp` `core/shader/recompiler/Translation/src/MemoryInstructions/MemoryResourceHelpers.cpp`
PR 14793	`docs/dev/debt/shader.md`
PR 14795	`docs/dev/debt/shader.md`
PR 14821	`core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp` `core/shader/recompiler/Translation/src/InstructionTranslator.cpp` `core/shader/recompiler/Translation/src/TranslationContext/OperandAccess.cpp` `docs/dev/debt/shader.md`
PR 14830	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp`
PR 14853	`core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 14871	`core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 14880	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/shader/recompiler/Recompiler.cpp`
PR 14888	`docs/dev/debt/shader.md`
PR 14892	`core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp` `core/shader/recompiler/Translation/src/TranslationContext/OperandAccess.cpp`
PR 14900	`docs/dev/debt/shader.md`
PR 14904	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `docs/dev/debt/shader.md`
PR 14909	`docs/dev/debt/shader.md`
PR 14913	`docs/dev/debt/shader.md`
PR 14944	`core/libs/prx/libSceAgcDriver/tests/Graphics.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`

klosixgaming-arch · 2026-10-19T00:52:28Z

Tested on Windows 11, MinGW-w64 GCC 15.2.0, NVIDIA GeForce RTX 5090 (driver 617.42), Ryzen 9 9950X3D. Head d662988bf merged onto current main (a373e85); the merge is clean.

agc_driver_wave_shared_vgprs, agc_driver_prepared_unregistered_graphics, agc_driver_graphics, agc_shader_memory, agc_shader_bindless and request_serializer all pass (6 of 6).

A note for other Windows testers: in a build folder under a long path (mine was a workspace path of about 150 characters) the build of agc_driver_prepared_unregistered_graphics_tests stopped with fatal error: opening dependency file ...PreparedUnregisteredGraphics.cpp.obj.d: No such file or directory. That is the Windows path-length limit, not this PR; the same tree built fine from a short build directory.

I did not run the new test against main without the fix, did not run PS5LM Plus (PPSA99581), and did not review the 28-file change (the HalfSwapU32 op, request version 33 and the cache key changes); this is a test report only, not an approval.

AI-assisted: yes. Written with AI assistance and reviewed by me.

wachi35676 · 2026-10-19T04:24:24Z

Rebased onto main 3ddbc4acc to resolve the conflict with #4650. #4650 took request serialization version 35 for its depth-range remap fields, so the shared VGPR fields now read from version 36. RequestSerializer writes 36 and accepts up to 36, and the test literals moved with it. The rest of the change is the same.

Windows 11 (WinLibs GCC 15.2.0, full build): the related subset (agc|shader|spirv|recomp|serial|wave) passes 779/779 on an RTX 5070 Ti. agc_driver_wave_shared_vgprs, request_serializer and agc_shader_memory also pass on the AMD RDNA2 iGPU (ANYPS5_GPU=AMD).
WSL Ubuntu 24.04 (GCC 13.3, llvmpipe): request_serializer and agc_shader_memory pass. The same 17 of 771 subset tests fail with and without this PR, because WSL lacks userfaultfd write protection ("require watched guest memory").

AI-assisted: yes (Claude Code); results above were run on this machine.

wachi35676 · 2026-10-19T06:50:13Z

Rebased again, onto main 46f27623c. The f32 flush-to-zero change took request serialization version 36, so the shared VGPR fields now read from version 37. TranslationContext and InstructionTranslator keep both setters. The rest of the change is the same.

Windows 11 (WinLibs GCC 15.2.0, full build), RTX 5070 Ti: the agc|shader|spirv|recomp|serial|wave subset passes 779 of 782.
- agc_driver_store_pending_identity passes 3 of 3 when run on its own.
- agc_driver_storage_file_write is intermittent: 1 of 3 here, and 1 of 3 on an unmodified main build.
- shader_cache_runtime_sources reaches its 180 s timeout with or without this PR while other builds load the machine.
AMD RDNA2 iGPU (ANYPS5_GPU=AMD): agc_driver_wave_shared_vgprs, request_serializer and agc_shader_memory pass.
WSL Ubuntu 24.04 (GCC 13.3, llvmpipe): request_serializer and agc_shader_memory pass. Failing here are the 17 that fail on main in WSL (no userfaultfd write protection), plus shader_cache_runtime_sources (a nested CMake configure exceeds 15 s under load). agc_driver_storage_file_write also failed in the full run and passes on its own.

AI-assisted: yes (Claude Code); results above were run on this machine.

blawesom · 2026-10-19T08:13:49Z

Rebase handoff — rebased onto current main cleanly (authorship unchanged).

Verified on Linux (CachyOS, kernel 7.2.9, GCC 16.2.1, Ninja, Release, BUILD_TESTING=ON, AMD RX 9070 XT / RADV): builds; conventions pass; full suite green apart from the four known main regressions (#5444 Mesa, #7584 BDA, #7766, #7784), none from this PR.

To pick it up:

git fetch https://github.com/blawesom/AnyPS5 rebase/7410
git checkout <your branch>
git merge --ff-only FETCH_HEAD

Head is 07b75d57b.

blawesom · 2026-10-19T08:19:15Z

One note on the handoff: an earlier full-suite run (on a stale build-tree state) showed agc_shader_memory and agc_driver_bda_span_reads failing — both pass on the rebuilt branch and on clean main (solo and in-suite), so those were stale-binary artifacts, not this PR. The handoff stands: head 07b75d57b, green apart from the four known main regressions (#5444 Mesa, #7584 BDA, #7766, #7784).

wachi35676 · 2026-10-19T08:24:13Z

@blawesom thanks for rebasing it and for the RX 9070 XT run. I compared 07b75d57b with the current head 39b3e272a, which was rebased onto 46f27623c earlier today: the added and removed lines of the patch are identical, so your base is the only difference. GitHub still reports this branch as mergeable, so I've left the head at 39b3e272a. I'll rebase again if a conflict appears.

AI-assisted: yes (Claude Code); results above were run on this machine.

LeandroLP23 · 2026-10-19T09:16:51Z

PS5LM Plus on an RDNA 3.5 iGPU: with this PR it gives the CPU backend's exact answer in all three configurations. Here the title isn't garbage on AMD.

AMD Radeon 860M (1002:1114, wave64 subgroups), driver 32.0.13058.8001, Windows 11, WinLibs GCC 15.2.0, Release. main f78675a, alone and with 39b3e27 merged. PS5LM Plus 0.8.4 (catalog zip, SHA-256 checked), Qwen2.5-0.5B-Instruct q4_k_m from Qwen's Hugging Face repo. app-args.txt holds the planner's arguments (-ngl 999 -c 4096 -b 256 -ub 64 -np 1 ...) with only the -fa/-ctk/-ctv part changed. Prompt "The capital of France is", n_predict 16, temperature 0, over HTTP on 8081. Three runs per cell, in the same run folder:

Configuration	`main`	This PR
`-fa on -ctk q8_0 -ctv q8_0`	`orntymology. In the 19th-arthury.10` (2 of 3; the first run stopped, see below)	`Paris. It is the largest city in Europe and the second largest in the world` (3 of 3)
`-fa on` (f16 K/V)	`Paris. It is the largest city in the world by population. It is also` (3 of 3)	`Paris. It is the largest city in Europe and the second largest in the world` (3 of 3)
`-fa off`	`Paris. It is the largest city in Europe and the second largest in the world` (2 of 3)	same (2 of 3)

This is the same pattern you measured on the RTX 5070 Ti, with one difference: on this GPU main's -fa off already gave the exact answer. The garbage you see on the RDNA 2 iGPU in every configuration doesn't happen here, so whatever causes it is specific to that GPU or its driver.

The new tests (agc_driver_wave_shared_vgprs, agc_driver_prepared_unregistered_graphics, agc_driver_graphics, agc_shader_memory, request_serializer) pass 3 of 3 on this head with ANYPS5_NO_SHADER_CACHE=1.

A separate stop, not from this PR. In three cases, the first run of a configuration on a fresh build stopped about 11 s after start, before the server answered: main q8_0, and -fa off on both trees. llama.cpp logs ggml_vulkan: device lost on Vulkan0, from waitForFences / waitSemaphores with a UINT64_MAX timeout, and AnyPS5 logs no error. The second and third runs of the same configuration, with the shader cache from the first, answered normally. That looks like the title's Vulkan driver giving up on a wait while AnyPS5 is still building pipelines, as in #6118, but I haven't traced it.

AI-assisted: yes (Claude Code)

DavidAmsellem · 2026-10-19T17:22:01Z

Second AMD-on-Linux (RADV) run, this time on RDNA3, after @blawesom's RX 9070 XT (RDNA4). Merged onto main dbac9fa15 (clean merge), built in full and compared with the same main, on the same machine.

AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 (Ubuntu 24.04 base), kernel 7.0.0-38, GCC 13.3, Release.

Full suite (ANYPS5_REQUIRE_VULKAN=1 ctest -j8): 1639/1641 with this PR, against 1638/1640 on main. The failures are the same two on both, so this PR adds none: agc_driver_image_table_vouched_lookups (#7766, fixed by #7916) and agc_driver_guest_memory_lifetime (#7784).

The new agc_driver_wave_shared_vgprs runs here and passes (its only skip is the subgroup-size check, which RADV's 64-lane subgroups clear), so RADV covers the test that llvmpipe skips. No other test is skipped or removed compared with main.

GPUs tested: AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 / kernel 7.0.0-38

AI-assisted: yes

DavidAmsellem · 2026-10-19T18:16:43Z

First AMD-on-Linux (RADV) run of this PR. Merged onto main dbac9fa15 (clean merge), built in full and compared with the same main, on the same machine.

AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 (Ubuntu 24.04 base), kernel 7.0.0-38, GCC 13.3, Release.

Full suite (ANYPS5_REQUIRE_VULKAN=1 ctest -j8) at 9033a683a: 1639/1641 with this PR, against 1638/1640 on main. The two failures are the same on both, so this PR adds none: agc_driver_image_table_vouched_lookups (#7766, fixed by #7916) and agc_driver_guest_memory_lifetime (#7784).

The PR's new ctest entry, agc_driver_wave_shared_vgprs, runs and passes here. No other ctest entry is added, skipped or removed compared with main.

This checks correctness only; I did not measure performance.

GPUs tested: AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 / kernel 7.0.0-38

AI-assisted: yes

wachi35676 · 2026-10-21T12:14:47Z

Rebased onto main 3e91c8edf, keeping @adriadam10's commit. main now writes request serialization version 42, so the shared VGPR fields read from version 43. The test literals moved with them, and the serializer commit's message lists the versions taken. Head 691c10187.

Windows 11, WinLibs GCC 15.2.0:

RTX 5070 Ti: the agc|shader|spirv|recomp|serial|wave subset passes 882 of 886. The four are:
- agc_driver_image_array_window_past_layer_limit (MinGW build break on main, test(agc): align the array window test's texels at run time #8310);
- three agc_driver_sample_order* tests, which fail identically in a build of main without this PR.
AMD RDNA2 iGPU: agc_driver_wave_shared_vgprs, request_serializer, agc_shader_memory and agc_driver_graphics pass, 15 of 15.

Thanks @DavidAmsellem for the RX 7900 XT RADV run.

AI-assisted: yes (Claude Code); results above were run on this machine.

wachi35676 · 2026-10-25T20:37:24Z

Rebased onto main 2f506ca (now a3fe509). main writes request serialization version 46 now: 42 sample order, 43 point list, 44 fused exact arithmetic, 45 pixel function LDS buffer, 46 target merge. This branch therefore moves to version 47, and the two wave-shared VGPR words are read only from 47.

I added a request_serializer case for a version 46 compute request. Without the version 47 gate the reader fails with RequestSerializer: truncated data.
I shortened the shader debt entry to below the new 1000-character limit (it now states the gap first). check_conventions.py passes.
Windows 11, WinLibs GCC 15.2: clean full build. The serializer, shader memory, wave-shared VGPR, dispatch, subgroup and shader cache set passes 24/24 on the RTX 5070 Ti and 24/24 on the AMD RDNA2 iGPU.
ctest -R ^agc_ on NVIDIA: 1050/1052. The two failures, agc_host_import_cut_lifetime (agc_host_import_cut_lifetime fails on Windows since the merge of #9159 #10746) and agc_driver_depth_plane_partial_store (AGC: imported depth-plane partial-store test fails refresh-proof assertion on Windows (NVIDIA and AMD) #10664), also fail on main.

AI-assisted: yes (Claude Code); results above were run on this machine.

wachi35676 · 2026-10-26T01:23:54Z

One more PS5LM Plus (PPSA99581) case this PR fixes, with a different model and a longer prompt: the flash-attention path with f16 K/V (the app's default -fa on -ctk f16 -ctv f16 -b 256 -ub 64), once the prompt is longer than a few tokens.

Model: SmolLM2-360M-Instruct q8_0 from HuggingFaceTB's GGUF repo (Apache-2.0), from /mnt/usb0/PS5LM/models. Requests go to the app's llama-server on port 8081 from the host: /completion, temperature 0, 24 tokens. Windows 11, Ryzen 5 9600X, RTX 5070 Ti (driver 621.01), WinLibs GCC 15.2.0, Release. These runs set APS5_SPANNED_WAVE=1, because without it the first use of the largest kernel (650,954 SPIR-V words, 11.3 s to build) trips the app's 10 s fence wait on a cold cache (radv/ps5: the GPU has not reached submission 561 (at 548) in 10 s, then ggml_vulkan: device lost on Vulkan0). That is a separate issue.

prompt	this PR's base `2f506caca`	this PR `a3fe509d5`
`The capital of France is` (5 tokens)	`Paris.\n\nParis is the capital of France. …`	same
`Write one sentence about the sea:`	`\n\nThe sea is a vast and mysterious place, full of secrets and wonders waiting to be discovered.`	same
`List three colors of the rainbow, separated by commas:` (11 tokens)	`side line line - - - - - - - - - …`	`\n\n* Blue\n* Green\n* Red\n\nYour response should contain at least 3 …`
the same request again (prompt cache hit, 1 token evaluated)	`...\n ...\n Shore... Shore... Shore... Zilililililil`	identical to the first answer

The PR column holds in 3 of 3 server runs, and the repeated request returns the same text each time.
The base column is one run on 2f506caca. The same garbage text comes on main aeff26bd9 in 3 of 3 runs: NVIDIA with spanning, NVIDIA without it (with a warm pipeline journal), and the AMD Radeon iGPU (gfx1036, ANYPS5_GPU=AMD).
On the AMD Radeon iGPU (gfx1036) with this PR, without spanning (64-lane subgroups), one run gives the same three answers as NVIDIA, and the repeated request matches the first. So with this model and configuration the iGPU no longer gives garbage; I haven't tried your Qwen2.5 q4_k_m configuration there.

AI-assisted: yes (Claude Code); results above were run on this machine.

… halves A wave64 compute program declares COMPUTE_PGM_RSRC3.SHARED_VGPR_CNT blocks of eight wave-shared VGPRs right after its (COMPUTE_PGM_RSRC1.VGPRS + 1) * 4 private ones (RDNA2 ISA 3.6.5): lane N and lane N + 32 read and write one copy. ACO swaps the halves of a wave64 through two of them on GFX10 (emit_permlane64_shared_vgpr) and emulates a wave-wide ds_bpermute_b32 with them (emit_bpermute_shared_vgpr). The flash-attention reductions of llama.cpp's Vulkan backend reach the swap. The recompiler kept these VGPRs private, so each half read its own stale copy and the upper half's partial sums were dropped. The driver now decodes both fields into ShaderComputeStageInfo (request serialization version 37, cache key and dispatch key included). A VALU write to a shared VGPR also updates the partner lane's copy through a new HalfSwapU32 exchange, the high half winning when both halves write, as a wave64 VALU instruction issues its low half first (ISA 4.2.1 Table 9, 6.2.4). A non-VALU write, such as ACO's ds_bpermute_b32 into a shared VGPR, is modeled while an immediate EXEC write earlier in the block has left one half of EXEC empty; with both halves possibly active the ISA gives no order and it throws, as do a VALU instruction that reads the shared VGPR it writes, v_writelane_b32, relative VGPR writes in such a program, and a draw whose SPI_SHADER_PGM_RSRC2.SHARED_VGPR_CNT is set. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ion version 50 main now writes request serialization version 49 (37 stencil operation value export, 38 mesh barycentric flag, 39 merged hull flag, 41 DenormPreserve 32, 42 sample order, 43 point list, 44 fused exact arithmetic, 45 pixel function LDS buffer, 46 target merge, 47 two-sample alpha-to-coverage dither, 48 fragment barycentric builtins, 49 depth export among them), so a compute request of version 49 or older written by main would be read with the two shared VGPR words it does not have. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JKJicKPXaPdKfXumezyTiE

wachi35676 · 2026-10-28T11:35:47Z

Rebased onto main 3a40945 (now 48fb6f0). main writes request serialization version 49 now: 47 two-sample alpha-to-coverage dither, 48 fragment barycentric builtins, 49 depth export. This branch moves to 50 and reads the two wave-shared VGPR words only from 50.

Resolved conflicts:
- StencilReplay.cpp: main's new ballot cases are kept, with HalfSwapU32 added to the opcode list next to BpermuteU32 as before.
- TranslationContext.hpp and InstructionTranslator.cpp: both setters are kept.
The compute replay case now uses a version 49 request. With the reader taking the two words from 37 as before, it fails with RequestSerializer: truncated data.
Windows 11, WinLibs GCC 15.2: clean full build. The serializer, shader-memory, wave-shared VGPR, dispatch, subgroup, masked-select, stencil and shader-cache set passes 40/40 on the RTX 5070 Ti and on the AMD RDNA2 iGPU. ctest -R ^agc_ on NVIDIA: 1122/1122. The only failure on the first run was agc_shader_memory checking for version 49; that check now expects 50.
check_conventions.py passes.

AI-assisted: yes (Claude Code); results above were run on this machine.

TechLevelZero · 2026-10-31T16:12:03Z

✅ NVIDIA GPU test: passed

PR head 48fb6f0 merged onto main 73f4f7e · tested 2026-10-31 16:12 UTC · 26 min


GPU	NVIDIA RTX 2000 Ada Generation, driver 595.99.02, Vulkan 1.4.329
Host	Ubuntu 26.04.1 LTS, kernel 7.0.0-38-generic, container ubuntu:24.04
Logs	build and test logs for this run
Build	✅ passed (Release, `-DANYPS5_ENABLE_SPIRV_TOOLS=ON`)
Tests	✅ 2,354 passed · 5 failed (0 new, 4 also fail on main, 1 unstable on main) · 48 skipped (`ANYPS5_REQUIRE_VULKAN=1`)
Test setup	CPU-only tests without a GPU; GPU tests (and any CPU-only test that failed or skipped without one) once on the GPU with `VK_LAYER_KHRONOS_validation` and `ANYPS5_NO_SHADER_CACHE=1`; main ran the same way
Test selection	2,407 of 3,072 tests run: the ones this change can reach, chosen by `tools/select_tests.py` from main's build graph
Vulkan validation	✅ no new validation IDs vs main (888 tests run on the GPU)
Slow tests (>30 s, GPU tests timed with validation layers)	⚠️ `agc_driver_bda_device` 55.5 s, ⚠️ `agc_driver_graphics_maps_fallback` 31.9 s, ⚠️ `agc_driver_graphics` 30.3 s

Unstable on main (1, these also failed in recent main runs on this machine; not counted against this PR)

agc_driver_submit_progress: failed in 7 of the last 10 main runs

Failing on main too (4, not counted against this PR)

agc_submit_progress
agc_submit_progress_no_timeline
guest_mmap_executable
guest_socket_options_multicast

Validated but not compared (1, not in the main baseline's validation pass)

agc_driver_wave_shared_vgprs

Validation IDs already present on main (3 IDs in 392 tests)

VUID-VkDeviceCreateInfo-pNext-pNext (392 tests): agc_compute_pipeline_replay, agc_driver, agc_driver_address_passes, agc_driver_address_passes_draw_transitions, agc_driver_address_passes_targets_split, agc_driver_alias_copy_label, agc_driver_alpha_to_coverage, agc_driver_array_layer_target, and 384 more
VUID-vkCmdCopyImage-srcImage-01548 (12 tests): agc_driver_depth_clear_hle, agc_driver_depth_plane_buffer_memory, agc_driver_depth_plane_cpu_write_after_memo, agc_driver_depth_plane_cpu_write_after_page_unwatch, agc_driver_depth_plane_cpu_write_after_partial_unwatch, agc_driver_depth_plane_partial_store, agc_driver_depth_plane_partial_store_unimported, agc_driver_depth_plane_sampling, and 4 more
VUID-vkCmdCopyImage-srcImage-01551 (12 tests): agc_driver_depth_clear_hle, agc_driver_depth_plane_buffer_memory, agc_driver_depth_plane_cpu_write_after_memo, agc_driver_depth_plane_cpu_write_after_page_unwatch, agc_driver_depth_plane_cpu_write_after_partial_unwatch, agc_driver_depth_plane_partial_store, agc_driver_depth_plane_partial_store_unimported, agc_driver_depth_plane_sampling, and 4 more

Skipped (48)

agc_driver_bda_reads_shadowed_results: skipped, the device reports watched host memory written when it imports or reads it
agc_driver_bda_reads_shadowed_results_per_build: skipped, the device reports watched host memory written when it imports or reads it
agc_driver_msaa_dcc_compared_import_2: skipped, the device reports host imports that keep write protection, so the write-watch probe compares none of them
agc_driver_msaa_dcc_compared_import_4: skipped, the device reports host imports that keep write protection, so the write-watch probe compares none of them
agc_driver_msaa_dcc_compared_import_8: skipped, the device reports host imports that keep write protection, so the write-watch probe compares none of them
agc_driver_sample_order_grid_reference: skipped, the device does not keep depth drawn at custom sample locations at those locations, even after a transition that names them: the composite of uniform renders holds the guest's depth in 0 of 3005 sloped samples, so no path can be compared with it bit for bit
agc_driver_vertex_elected_atomic_max: skipped, the device runs vertex programs in subgroups, so they keep their cross-lane reductions
agc_host_import_sparse: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_import_settle_initial_dcc: skipped, the device has no deferred settle of host imports: imports under 1 MiB wait for one only on Windows
agc_unit_shadow_completion_store: skipped, the device reports watched host memory written when it imports or reads it
guest_http_uri_locale: 
guest_net_ctl: skipped, the host has no IPv4 address
guest_video_out_color_space: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_hdr: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_high_frame_rate: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_output_mode: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_resolution_status: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_wqhd_detection: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
video_out_flip_aftergpu: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_backtoback: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_controls: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_cpuflip: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_cpuflip_in_order: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_dcc: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_deferred: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_eop: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or plat

Reproduce

git fetch origin 48fb6f0c67ab3e0d80cc1eb820448aff27697050 73f4f7e06e4f210a8c6570e335fd654dc0a86134
git checkout 73f4f7e06e4f210a8c6570e335fd654dc0a86134 && git merge --no-ff 48fb6f0c67ab3e0d80cc1eb820448aff27697050
git submodule update --init --recursive
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON -DANYPS5_ENABLE_SPIRV_TOOLS=ON
cmake --build build && cmake --build build --target libs
VK_DRIVER_FILES=/nonexistent/icd.json ctest --test-dir build -E '^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu' --timeout 120 --parallel 32 --output-on-failure
ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation \
  ctest --test-dir build -R '^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu' --timeout 120 --parallel 4 --output-on-failure
# failures are then re-run alone: --timeout 360 --parallel 2; new validation IDs: the test alone, twice, --parallel 1

_{Automated build and test on a contributor's NVIDIA machine; it does not replace upstream CI. A new comment is posted for each new push. Full logs.}

github-actions Bot added shader core/shader docs agc libSceAgc* unblocks title Names a title ID on the Unblocks line needs rebase Conflicts with main, rebase needed labels Oct 18, 2026

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from 70e15b4 to d662988 Compare October 18, 2026 23:16

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 18, 2026

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from d662988 to e68ea55 Compare October 19, 2026 01:43

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

SleglJan mentioned this pull request Oct 19, 2026

feat(agc): run registered merged hull programs (binary type 3) #6679

Merged

11 tasks done

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from e68ea55 to 455d68f Compare October 19, 2026 04:24

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from 455d68f to 39b3e27 Compare October 19, 2026 06:49

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

adriadam10 self-assigned this Oct 19, 2026

adriadam10 force-pushed the fix/shader-wave-shared-vgprs branch from 39b3e27 to 5cea227 Compare October 19, 2026 16:13

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Oct 19, 2026

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from a165137 to 691c101 Compare October 21, 2026 12:14

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 21, 2026

imellag mentioned this pull request Oct 21, 2026

feat(recompiler): read I/J from barycentric built-ins on MoltenVK #9193

Merged

12 tasks done

akkadbakkad89 mentioned this pull request Oct 21, 2026

fix(recompiler): exempt proven control conditions and own-bit ballots in stencil replay #9317

Merged

12 tasks done

MarcZX10 mentioned this pull request Oct 22, 2026

fix(recompiler): let optional-body merges reach vetted ballots and wave branches in stencil replay #9394

Closed

12 tasks done

EagleBearerYS mentioned this pull request Oct 22, 2026

fix(shader): print memory instructions' pc and offset in the IR dump #9907

Merged

12 tasks done

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Oct 23, 2026

Filip722 mentioned this pull request Oct 23, 2026

perf(shader): skip log2 correction outside the near-one range #10336

Merged

12 tasks done

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from 691c101 to a3fe509 Compare October 25, 2026 20:36

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 25, 2026

QRMOE mentioned this pull request Oct 25, 2026

fix(shader): keep pixel helper lanes out of guest lane masks #11199

Merged

12 tasks done

wachi35676 mentioned this pull request Oct 26, 2026

perf(shader): span single-wave wave64 compute by default after a uniformity check #7922

Merged

12 tasks done

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Oct 26, 2026

milescoviello mentioned this pull request Oct 27, 2026

perf(shader): look up opcode access traits in a table built at compile time #12002

Draft

12 tasks done

wachi35676 and others added 2 commits October 28, 2026 16:14

wachi35676 force-pushed the fix/shader-wave-shared-vgprs branch from a3fe509 to 48fb6f0 Compare October 28, 2026 11:35

github-actions Bot removed the needs rebase Conflicts with main, rebase needed label Oct 28, 2026

frankea mentioned this pull request Oct 31, 2026

fix(shader): run tessellation evaluation stages one lane per invocation on devices without subgroups there #13807

Merged

12 tasks done

github-actions Bot added the needs rebase Conflicts with main, rebase needed label Nov 1, 2026

fogydes mentioned this pull request Nov 1, 2026

perf(shader): pick the f32 NaN result with one quiet and one fallback select #4287

Open

9 tasks done

This was referenced Nov 3, 2026

perf(shader): stop zeroing mesh window buffer slots at workgroup start #15362

Merged

perf(shader): keep the vector window to the blocks between relative accesses #15520

Open

LeandroLP23 mentioned this pull request Nov 5, 2026

AGC: PS5 FSR4 Showcase (PPSA99011) gives up in its first FSR4 pass while the queue worker builds 130k-800k-word compute pipelines (up to 66 s each on AMD) #6118

Open

Filip722 mentioned this pull request Nov 6, 2026

perf(shader): evaluate only the selected cosine polynomial #16350

Open

12 tasks done

Conversation

wachi35676 commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Tested

Checklist

Uh oh!

github-actions Bot commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

klosixgaming-arch commented Oct 19, 2026

Uh oh!

wachi35676 commented Oct 19, 2026

Uh oh!

wachi35676 commented Oct 19, 2026

Uh oh!

blawesom commented Oct 19, 2026

Uh oh!

blawesom commented Oct 19, 2026

Uh oh!

wachi35676 commented Oct 19, 2026

Uh oh!

LeandroLP23 commented Oct 19, 2026

Uh oh!

DavidAmsellem commented Oct 19, 2026

Uh oh!

DavidAmsellem commented Oct 19, 2026

Uh oh!

wachi35676 commented Oct 21, 2026

Uh oh!

wachi35676 commented Oct 25, 2026

Uh oh!

wachi35676 commented Oct 26, 2026

Uh oh!

wachi35676 commented Oct 28, 2026

Uh oh!

TechLevelZero commented Oct 31, 2026

✅ NVIDIA GPU test: passed

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

wachi35676 commented Oct 18, 2026 •

edited

Loading

github-actions Bot commented Oct 18, 2026 •

edited

Loading