fix(shader): widen LDS atomic fences to device scope on AMD's proprietary driver by wachi35676 · Pull Request #8489 · boykopovar/AnyPS5

wachi35676 · 2026-10-20T13:41:29Z

Builds on #8134, which has merged; this branch is now a single commit on main.

What

The recompiler puts a workgroup-scope OpMemoryBarrier(Workgroup, AcquireRelease | WorkgroupMemory) after three kinds of LDS update:

every LDS atomic;
the compare-exchange loop that writes a byte or a halfword of LDS (ds_write_b8/ds_write_b16);
misaligned LDS stores.

AMD's proprietary Windows driver compiles that memory barrier to a real s_barrier. I read the driver's code objects out of its pipeline cache to see this; capturing VK_KHR_pipeline_executable_properties internal representations fails with -13 for this pipeline.

These updates usually sit in branches that only some waves of a workgroup take. Those waves then execute extra s_barriers, which pair with the other waves' real barriers, so the other waves read LDS before it is written.

An OpMemoryBarrier is not an execution barrier, and SPIR-V allows it in non-uniform control flow. The SPIR-V is valid, so this is a driver bug. On the same driver, a device-scope OpMemoryBarrier compiles to buffer_gl0_inv + buffer_gl1_inv with no s_barrier.

PS5LM Plus (PPSA99581, llama.cpp's Vulkan backend) drifted into different garbage on every run on that driver with flash attention on. Its flash-attention shader stores two halfwords to LDS inside the KV loop, in a branch that waves 0-1 of its four take. Results from replaying that dispatch with identical inputs:

LDS fence on AMD	`s_barrier` in the driver's code (guest has 21)	Replay vs NVIDIA (0x7380 bytes of output)
Workgroup scope (before)	23	159 words differ, up to 1.7e38
Device scope (this PR)	21	9 words differ in the last bit (≤ 3e-8)
Subgroup scope / no fence (experiments only)	21	the same 9 last-bit differences

The guest's 21 s_barriers map to 21 OpControlBarriers, which the driver compiles to one s_barrier each, none moved or duplicated. The 2 extra come from the two halfword stores.

Changes:

WorkgroupScopeFenceUsable(VkDriverId) in SubgroupBarrier.hpp is false only for VK_DRIVER_ID_AMD_PROPRIETARY. NVIDIA, RADV, AMDVLK, Intel, llvmpipe and MoltenVK keep the workgroup scope. AMDVLK is excluded only because it was not tested.
The device passes that as the new SpirvTarget::deviceScopeLdsMemoryBarriers, which is part of the shader cache key.
With the flag set, the LDS fence (LdsMemoryBarrierScope) is device scope, with the same semantics. Device scope includes the workgroup, so it orders everything the old fence did. Modules use the GLSL450 memory model, and device-scope fences are already emitted for buffer atomics.
SharedAtomicUpdate (ds_min/max_f32, ds_cmpst_f32 and the other read-modify-write LDS operations) emitted a second fence identical to the one AtomicUpdate already emits. The duplicate is dropped, so each update has one fence on every driver.
agc_shader_replay --device-scope-lds-memory-barriers replays a request the way that driver compiles it. Like --wave-lds-memory-barriers, the flag is not serialized.
One debt line in docs/dev/debt/shader.md.

The fences inside the LDS lock path are untouched, because that path emits none. On that driver a ds_add_u64 program (the lock path) had no s_barrier besides its uniform lock-initialisation barrier, and an LDS exchange after it in a branch passed.

Not covered:

Only one GPU and driver was measured: the Radeon iGPU of a Ryzen 5 9600X (0x13c0, RDNA 2), Windows driver 32.0.21042.62.
LeandroLP23 reports that a Radeon 860M (RDNA 3.5, 32.0.13058.8001) gives correct flash-attention output even without fix(shader): order wave LDS accesses with a memory barrier on AMD's proprietary driver #8134. The gate still covers every AMD proprietary version.
The cost of the wider fence on AMD (GL0 and GL1 invalidation per LDS update) was not timed separately. PS5LM Plus decodes at the same speed with and without it (2-4 tokens/s on that iGPU).

Tested

Windows 11, WinLibs GCC 15.2, full Release build of 98f4019 (on #8134 473497c, main 3aa35dc; the branch also merges cleanly with main 44e3389). ctest -R "agc|shader|spirv|recomp" ran with ANYPS5_REQUIRE_VULKAN=1 and -j4; failures were rerun alone. The baseline is #8134's head 473497c.

NVIDIA GeForce RTX 5070 Ti (617.42): 827 tests, 2 failed: agc_driver_msaa_dcc_clear_4 and agc_driver_pass_input_hazard. The baseline has the same 2 of 826.
AMD Radeon Graphics (0x13c0, 32.0.21042.62): 827 tests, 5 failed:
- agc_trimmed_import, agc_driver_msaa_dcc_clear_4, agc_driver_draw_cache_reuse_off and agc_driver_stencil_replay_no_gpl also fail on the baseline.
- agc_driver_layered_volume_vertex_no_gpl fails with a device fault reading 0xbeefdeadc000, a fault the debt file already lists for this driver. The family's failures depend on test order: on the baseline the same fault hits agc_driver_layered_volume_draw and agc_driver_layered_volume_vertex instead (baseline: 8 of 826 failed).

Linux (WSL Ubuntu 24.04, GCC 13.3, llvmpipe): full build with -DBUILD_TESTING=ON, the same ctest set, 817 tests. The same 17 tests fail as on #8134; they need watched guest memory or llvmpipe features. agc_subgroup_barrier and agc_shader_disk_cache pass, and agc_driver_lds_subword_store_branch skips (no 32-lane subgroups).

New and changed tests:

agc_driver_lds_subword_store_branch runs four wave32 waves. Waves 0-1 store a halfword to LDS in a branch and then spin for 4096 iterations. After that, every thread writes an LDS slot, passes one s_barrier and reads the slot of thread tid ^ 64. The test runs 16 rounds.
- With the scope widening disabled (negative control), it fails 3 of 3 times on AMD: round 0: thread 64 (wave 2) read 868449278, expected 65536.
- With the change it passes 3 of 3 on AMD and on NVIDIA. NVIDIA passes either way.
agc_subgroup_barrier also checks the new quirk, and that ds_write_b16, ds_add_u32 and ds_max_f32 each emit exactly one fence of the expected scope with the flag off and on. It fails without the change.
agc_shader_disk_cache checks that the new flag changes the cache key.

PS5LM Plus v0.8.4 was run with qwen2.5-0.5b-instruct-q4_k_m.gguf, -fa on -ngl 999, the prompt "The capital of France is" and temperature 0, on the AMD iGPU with a warmed shader cache per configuration. Flash attention also needs #7410: without it, NVIDIA too answers garbage with -fa on. So these runs use a local integration branch, this PR (#8134 + this commit) merged with #7410 (a80f17d). The "before" rows are the same build with the new flag forced off.

KV cache	Before (flag off)	After
q8_0	4 runs, 4 different texts, e.g. `" Paris. It is not a city and not to\",__ F 12G"`	3 of 3: `" Paris. It is the largest city in Europe and the second largest in the world"`
f16	4 runs, 4 different texts, e.g. `" Paris. It is in France, in France and or in France, in France"`	4 of 4: the same text as q8_0

The correct text matches NVIDIA and the CPU. The first cold q8_0 run of the final commit hit the guest's own 10 s fence timeout while shaders compiled. The previous revision of this commit, with the duplicate fence still in place, gave the same correct text on 5 q8_0 and 4 f16 runs.

Checklist

🤖 Generated with Claude Code

https://claude.ai/code/session_01BM2kfDbr6MydxsfTckZJDR

github-actions · 2026-10-20T13:55:35Z

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on `main`)	Same exports	Same new files
PR 6959	`docs/dev/debt/shader.md` L191

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/Execution/src/VulkanDevice.cpp`
PR 2434	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 4040	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 4072	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/Recompiler.cpp`
PR 4472	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 5021	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvMemoryInstructions.cpp` `docs/dev/debt/shader.md`
PR 5371	`docs/dev/debt/shader.md`
PR 5431	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 5441	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 5609	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 5925	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6188	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6202	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 6235	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 6280	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `docs/dev/debt/shader.md`
PR 6412	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 6488	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `docs/dev/debt/shader.md`
PR 6539	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6858	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp`
PR 6924	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 6960	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 6977	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 7085	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `docs/dev/debt/shader.md`
PR 7201	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp`
PR 7275	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 7307	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 7369	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 7449	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 7484	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `docs/dev/debt/shader.md`
PR 7550	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp`
PR 7618	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 7642	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvMemory/SpirvBufferAccess.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvBufferAccess.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvMemoryInstructions.cpp`
PR 7677	`core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp`
PR 7684	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `docs/dev/debt/shader.md`
PR 7689	`core/shader/recompiler/Recompiler.hpp`
PR 7922	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvMemoryInstructions.cpp` `docs/dev/debt/shader.md`
PR 8004	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvMemoryInstructions.cpp` `docs/dev/debt/shader.md`
PR 8014	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 8077	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `docs/dev/debt/shader.md`
PR 9020	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `docs/dev/debt/shader.md`
PR 9049	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp`
PR 9183	`docs/dev/debt/shader.md`
PR 9193	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 9335	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `docs/dev/debt/shader.md`
PR 9348	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 9356	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp`
PR 9387	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp`
PR 9400	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 9431	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9432	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9508	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvMemory/SpirvBufferAccess.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvBufferAccess.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvMemoryInstructions.cpp`
PR 9633	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9672	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9673	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9674	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9675	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9676	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9748	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `docs/dev/debt/shader.md`
PR 9749	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9842	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9852	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 9889	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `docs/dev/debt/shader.md`
PR 9891	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9892	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9893	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9895	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/libs/prx/libSceAgcDriver/tools/ShaderReplay.cpp`
PR 9897	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 9898	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 9969	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.cpp` `docs/dev/debt/shader.md`
PR 9989	`core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 10000	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/shader.md`
PR 10049	`core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 10193	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp` `core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitter.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterHelpers.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvEmitterState.hpp` `core/shader/recompiler/SpirvBackend/include/SpirvBackend/SpirvMemory/SpirvBufferAccess.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvBufferAccess.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvMemoryInstructions.cpp`
PR 10194	`core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp`
PR 10348	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.hpp`
PR 10359	`core/shader/recompiler/CacheKey.hpp` `core/shader/recompiler/Recompiler.cpp` `core/shader/recompiler/Recompiler.hpp`

xXPBSXx · 2026-10-20T23:07:00Z

Independent Windows run of this PR on an RDNA 2 card (data only).

Environment: Windows 10 Pro 10.0.19045, Ryzen 7 5800X3D, AMD Radeon RX 6400 (AMD driver 32.0.21033.3005; the log shows Selected GPU name=AMD Radeon RX 6400 vendor=0x1002 device=0x743f subgroup=64, picked with ANYPS5_GPU="RX 6400"), WinLibs GCC 15.2.0 (MinGW-w64 UCRT), Release. PR head c1528a6e18; merge-tree against main is clean. I built agc_driver_lds_subword_store_branch_tests, agc_driver_lds_tree_reduction_tests and agc_subgroup_barrier_tests.

Results (3 runs each):

PR head: agc_driver_lds_subword_store_branch, agc_driver_lds_tree_reduction and agc_subgroup_barrier pass 3/3.
Baseline: I kept the PR's tests and changed only WorkgroupScopeFenceUsable in SubgroupBarrier.hpp to return true;, so AMD's proprietary driver keeps the workgroup-scope fence again, then rebuilt. agc_driver_lds_subword_store_branch fails 3/3 on this card with thread 64 (wave 2) read <stale value>, expected 65536 (the stale value was 1135280128 in two runs and 1065353216 in one); agc_subgroup_barrier fails 3/3 (driver 1: expected workgroup-scope fences to be unusable, a check of the quirk itself); agc_driver_lds_tree_reduction passes 3/3 in the baseline too, so it does not distinguish the two states on this card. I restored the PR's header afterwards.

So the stale LDS read the PR describes reproduces on an RX 6400 without the device-scope fence and is gone with it. I did not inspect the driver's code objects, so I can not confirm the s_barrier explanation, and my driver version differs from the one in the PR text.

Not run: PS5LM Plus or any title, the replay comparison against NVIDIA, other GPUs and platforms, the full repository CTest.

AI-assisted (Claude Code).

wachi35676 · 2026-10-21T12:14:45Z

Rebased onto main 3e91c8edf now that #8134 has merged, so this is a single commit: f1a210313. The conflicts were in CacheKey.hpp, Recompiler.hpp and SpirvEmitter.cpp, where main added float32DenormPreserve beside this PR's deviceScopeLdsMemoryBarriers; both are kept. The code is otherwise unchanged.

Windows 11, WinLibs GCC 15.2.0, agc|shader|spirv|recomp|subgroup subset:

RTX 5070 Ti: 881 of 885 pass. The four are:
- agc_driver_image_array_window_past_layer_limit, which doesn't compile with MinGW on main (test(agc): align the array window test's texels at run time #8310 fixes it);
- three agc_driver_sample_order* tests ("the slopes did not leave depth both at and below the clear value"), which this PR doesn't touch on NVIDIA. They fail the same way in a main + fix(shader): share wave64 VGPRs past the private ones between the two halves #7410 build.
AMD RDNA2 iGPU (ANYPS5_GPU=AMD): the same four, plus agc_driver_tessellation. That one is the merged hull failure on this driver that I reported on feat(agc): run registered merged hull programs (binary type 3) #6679; it also fails without this PR. agc_driver_lds_subword_store_branch, agc_driver_lds_tree_reduction and agc_subgroup_barrier pass.

Thanks @xXPBSXx for the RX 6400 run.

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

…tary driver The recompiler orders an LDS atomic, and the compare-exchange loop that writes a byte or a halfword of LDS (ds_write_b8/b16), with a workgroup-scope OpMemoryBarrier after the update. AMD's proprietary Windows driver 32.0.21042.62 (Radeon 0x13c0, RDNA 2) compiles that memory barrier to s_barrier, as the code objects in its pipeline cache show. Such updates usually sit in branches that only some waves of a workgroup take, so those waves execute extra s_barriers, which pair with the other waves' barriers, and the other waves read LDS before it is written. An OpMemoryBarrier is not an execution barrier and SPIR-V allows it in non-uniform control flow, so the SPIR-V is valid; a device-scope OpMemoryBarrier compiles to GL0 and GL1 cache invalidations without s_barrier there. PS5LM Plus (PPSA99581, llama.cpp's Vulkan backend) drifted into garbage on that driver with flash attention on. Its flash-attention shader stores two halfwords to LDS in a branch that two of its four waves take, inside the KV loop. The driver's code had 23 s_barrier for the guest's 21; a replay of the dispatch differed from NVIDIA in 159 words, up to 1.7e38, and now has 21 and differs only in the last bit of 9 words. On that driver the fence is now device scope, which orders everything the workgroup-scope fence did. Other drivers keep the workgroup scope. The flag is part of the shader cache key, and agc_shader_replay takes --device-scope-lds-memory-barriers to replay a request the way that driver compiles it. The LDS update helper already fences, so the second, identical fence after it in ds_min/max_f32 and the other read-modify-write LDS operations is dropped. agc_driver_lds_subword_store_branch stores a halfword to LDS in a branch that waves 0-1 of four take, delays those waves and then exchanges LDS slots across one s_barrier. Before, thread 64 read a stale slot on the AMD driver; it passes on AMD and NVIDIA now. agc_subgroup_barrier checks the driver quirk and that ds_write_b16, ds_add_u32 and ds_max_f32 each emit one fence of the expected scope. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BM2kfDbr6MydxsfTckZJDR

adriadam10 · 2026-10-23T22:04:04Z

Checked the diff, the new tests and the third-party AMD run. The fence scope only changes for VK_DRIVER_ID_AMD_PROPRIETARY, it is part of the shader cache key, and agc_subgroup_barrier pins one fence of the expected scope for ds_write_b16, ds_add_u32 and ds_max_f32. Rebased onto current main (adjacent fields in Recompiler.hpp kept beside the new one); merged.

github-actions Bot added the unblocks title Names a title ID on the Unblocks line label Oct 20, 2026

wachi35676 force-pushed the fix/shader-amd-lds-fence-scope branch from 98f4019 to c1528a6 Compare October 20, 2026 16:00

wachi35676 mentioned this pull request Oct 20, 2026

fix(shader): order wave LDS accesses with a memory barrier on AMD's proprietary driver #8134

Merged

12 tasks done

github-actions Bot added build shader core/shader docs agc libSceAgc* labels Oct 20, 2026

imellag mentioned this pull request Oct 20, 2026

perf(shader): emit exact f32 add, subtract and multiply as fma on MoltenVK #8612

Merged

12 tasks done

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

wachi35676 force-pushed the fix/shader-amd-lds-fence-scope branch from c1528a6 to f1a2103 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

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

adriadam10 self-assigned this Oct 23, 2026

adriadam10 force-pushed the fix/shader-amd-lds-fence-scope branch from f1a2103 to 98df163 Compare October 23, 2026 19:11

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

adriadam10 merged commit e7f46f5 into boykopovar:main Oct 23, 2026
10 checks passed

maximilianfv mentioned this pull request Oct 25, 2026

fix(agc): order wave LDS accesses with memory barriers on MoltenVK #10858

Merged

12 tasks done

Navigation Menu

Repository navigation

fix(shader): widen LDS atomic fences to device scope on AMD's proprietary driver - #8489

fix(shader): widen LDS atomic fences to device scope on AMD's proprietary driver#8489
adriadam10 merged 1 commit into
boykopovar:mainfrom
wachi35676:fix/shader-amd-lds-fence-scope

wachi35676 commented Oct 20, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 20, 2026 •

edited

Loading

Uh oh!

xXPBSXx commented Oct 20, 2026

Uh oh!

wachi35676 commented Oct 21, 2026

Uh oh!

Uh oh!

adriadam10 commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

wachi35676 commented Oct 20, 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 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

xXPBSXx commented Oct 20, 2026

Uh oh!

wachi35676 commented Oct 21, 2026

Uh oh!

Uh oh!

adriadam10 commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

wachi35676 commented Oct 20, 2026 •

edited

Loading

github-actions Bot commented Oct 20, 2026 •

edited

Loading