gpuav: Support coopmat in shared memory data race pass. by jeffbolznv · Pull Request #11780 · KhronosGroup/Vulkan-ValidationLayers

jeffbolznv · 2026-03-02T21:49:57Z

No description provided.

ci-tester-lunarg · 2026-03-02T21:50:01Z

CI Vulkan-ValidationLayers build queued with queue ID 666500.

ci-tester-lunarg · 2026-03-02T21:51:20Z

CI Vulkan-ValidationLayers build queued with queue ID 666513.

ci-tester-lunarg · 2026-03-02T21:55:03Z

CI Vulkan-ValidationLayers build # 22645 running.

ci-tester-lunarg · 2026-03-02T22:29:15Z

CI Vulkan-ValidationLayers build queued with queue ID 666602.

ci-tester-lunarg · 2026-03-02T22:30:01Z

CI Vulkan-ValidationLayers build # 22647 running.

ci-tester-lunarg · 2026-03-03T00:33:51Z

CI Vulkan-ValidationLayers build # 22647 failed.

ci-tester-lunarg · 2026-03-03T01:24:10Z

CI Vulkan-ValidationLayers build queued with queue ID 666699.

ci-tester-lunarg · 2026-03-03T01:25:01Z

CI Vulkan-ValidationLayers build # 22648 running.

spencer-lunarg · 2026-03-03T03:01:57Z

+        if ((ret & (STORE_BIT | ATOMIC_BIT)) != 0) {
+            error_payload = ErrorPayload(
+                            inst_offset,
+                            SpecConstantLinkShaderId | (kErrorGroup_SharedMemoryDataRace << kErrorGroup_Shift) | (kErrorSubCode_SharedMemoryDataRace_RaceOnLoad << kErrorSubCode_Shift),


happy to do in a follow up PR, but I think this should be a different error message to print it more clearly the error didn't have from a load, but a coop-load

The message will have the OpCooperativeMatrixLoadKHR opcode in it, so it's fairly clear.

spencer-lunarg · 2026-03-03T03:08:14Z

        return false;
    }

+    // relies on some subgroup functionality


... I don't think we need this, if the user doesn't have coop-mat, it will not link in any subgroup sutff

also SPV_KHR_cooperative_matrix relies on SPIR-V 1.3, so you couldn't have a coop-mat without out anyway

I did this to fix PositiveGpuAVBufferDeviceAddress.AtomicsWorkgroups which was failing in the previous CI run. Now that the sharedmemdatarace shader uses subgroup stuff, the inputs like SubgroupId get linked in, but the groupnonuniform capability is removed here:

// Vulkan 1.1 is required, so if incoming SPIR-V is 1.0, might need to adjust it const uint32_t spirv_version_1_0 = 0x00010000; if (header_.version == spirv_version_1_0) { // SPV_KHR_storage_buffer_storage_class is needed, but glslang removes it from linking functions AddExtension("SPV_KHR_storage_buffer_storage_class"); // Subgroups where added in Vulkan 1.1, so SPIR-V 1.0 can't use them // This is a bad hack around for someone using a SPIR-V 1.0 RemoveCapability(spv::CapabilityGroupNonUniform); }

It looks like the new failure is something different, so I guess this fix worked (I saw spirv-val failures, not sure how that is configured in CI)

... I've debated to just enforce 1.3 SPIR-V.... not going to worry too much about this for now

spencer-lunarg

some comments/idea, but can be applied in a follow-up PR if needed

ci-tester-lunarg · 2026-03-03T03:19:22Z

CI Vulkan-ValidationLayers build # 22648 failed.

jeffbolznv · 2026-03-03T03:34:28Z

I can't reproduce the current CI failure locally (NegativeGpuAVSharedMemoryDataRace.CoopMatStoreLoad). Is there anything useful in the logs?

spencer-lunarg · 2026-03-03T03:45:47Z

Note: Google Test filter = NegativeGpuAVSharedMemoryDataRace.CoopMatStoreLoad
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from NegativeGpuAVSharedMemoryDataRace
[ RUN      ] NegativeGpuAVSharedMemoryDataRace.CoopMatStoreLoad
Driver Name = NVIDIA
Driver Info = 550.78
/home/lunarg/.jenkins/vz3/Debug64/Vulkan-ValidationLayers/tests/framework/error_monitor.cpp:309: Failure
Failed
Validation Error: [ SharedMemoryDataRace-RaceOnLoad ] | MessageID = 0xf32d0b17
vkCmdDispatch(): A data race was detected on the shared memory variable "arr" in local invocation index 1 while performing a load operation. (Likely against local invocation index 0)
Stage = Compute.  Global invocation ID (x, y, z) = (1, 0, 0)
Command buffer (0x5bbba7df6e20)
	Compute Dispatch Index 0
Shader Module (0x140000000014) (internal ID 1)
SPIR-V Instruction: %42 = OpLoad %6 %41 48 19
(Unable to find shader source, build shader with debug info to get source information)

Objects: 3
    [0] VkQueue 0x5bbba6f81a90
    [1] VkCommandBuffer 0x5bbba7df6e20
    [2] VkPipeline 0x1b000000001b



[  FAILED  ] NegativeGpuAVSharedMemoryDataRace.CoopMatStoreLoad (1579 ms)
[----------] 1 test from NegativeGpuAVSharedMemoryDataRace (1579 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test suite ran. (1851 ms total)
[  PASSED  ] 0 tests.
[  FAILED  ] 1 test, listed below:
[  FAILED  ] NegativeGpuAVSharedMemoryDataRace.CoopMatStoreLoad

 1 FAILED TEST

... not sure what is going on, should have reported the SharedMemoryDataRace only once due to the limit

not sure what is going on either... worse case, can silent it until I have access to the machine to look more

jeffbolznv · 2026-03-03T05:00:45Z

I added printfs and was lucky enough to see it happen once. There can intermittently be both RaceOnStore and RaceOnLoad, and the duplicate message limit applies per-VUID.

[ RUN      ] NegativeGpuAVSharedMemoryDataRace.CoopMatStoreLoad
vuid_text VALIDATION-SETTINGS
vuid_text WARNING-CreateInstance-status-message
vuid_text WARNING-CreateInstance-debug-warning
vuid_text WARNING-Setting-Limit-Adjusted
vuid_text WARNING-Setting-Limit-Adjusted
vuid_text SharedMemoryDataRace-RaceOnStore
vuid_text SharedMemoryDataRace-RaceOnLoad
C:\github\jeffbolznv\Vulkan-ValidationLayers\tests\framework\error_monitor.cpp(309): error: Failed
Validation Error: [ SharedMemoryDataRace-RaceOnLoad ] | MessageID = 0xf32d0b17
vkCmdDispatch(): A data race was detected on the shared memory variable "arr" in local invocation index 1 while performing a load operation. (Likely against local invocation index 0)
Stage = Compute.  Global invocation ID (x, y, z) = (1, 0, 0)
Command buffer (0x23d005870d0)
        Compute Dispatch Index 0
Shader Module (0x140000000014) (internal ID 1)
SPIR-V Instruction: %42 = OpLoad %6 %41 48 19
(Unable to find shader source, build shader with debug info to get source information)

Objects: 3
    [0] VkQueue 0x23d004214d0
    [1] VkCommandBuffer 0x23d005870d0
    [2] VkPipeline 0x1b000000001b



vuid_text SharedMemoryDataRace-RaceOnLoad
vuid_text SharedMemoryDataRace-RaceOnLoad
vuid_text SharedMemoryDataRace-RaceOnLoad
vuid_text SharedMemoryDataRace-RaceOnLoad

Any clever ideas what to do about this? Seems like it could affect some other tests, too.

ci-tester-lunarg · 2026-03-03T06:32:44Z

CI Vulkan-ValidationLayers build queued with queue ID 666868.

jeffbolznv · 2026-03-03T06:33:09Z

I changed this and a few other tests to have a single race. This uncovered another bug, so I guess that worked out well.

ci-tester-lunarg · 2026-03-03T06:35:01Z

CI Vulkan-ValidationLayers build # 22649 running.

ci-tester-lunarg · 2026-03-03T07:53:08Z

CI Vulkan-ValidationLayers build # 22649 passed.

jeffbolznv requested a review from a team as a code owner March 2, 2026 21:49

gpuav: Support coopmat in shared memory data race pass

1715307

jeffbolznv force-pushed the shmem_coopmat branch from 76aed02 to 1715307 Compare March 2, 2026 21:51