Environment:
- OS: Ubuntu 26.04 LTS
- GPU and driver version: NVIDIA GeForce RTX™ 4070 SUPER - NVIDIA 595.71.05
- SDK or header version if building from repo: 1.4.350.1
- Options enabled (synchronization, best practices, etc.): 'GPU assisted validation'
Describe the Issue
The Vulkan Validation Layers trigger a false positive SharedMemoryDataRace-RaceOnAtomic error when non-atomic initialization of a groupshared array variable is followed by a subgroup-scoped control barrier (Slang function : GroupMemoryBarrierWithWaveSync) before subsequent InterlockedAdd operations within the same subgroup.
Even though the execution and memory visibility are properly synchronized at the Subgroup scope prior to atomic execution, the validation layer appears to require a full Workgroup-scoped execution barrier to clear the data race tracking state for Workgroup memory accesses.
Here the minimal code :
#define N_WAVES 256 / 32 // here the nvidia wave size is used
groupshared uint sharedBins[16 * N_WAVES];
[numthreads(256, 1, 1)]
void main()
{
uint waveBinsIndex = WaveGetWaveIndex() * 16;
if (WaveGetLaneIndex() < 16)
{
sharedBins[waveBinsIndex + WaveGetLaneIndex()] = 0;
}
//GroupMemoryBarrierWithGroupSync(); // works
GroupMemoryBarrierWithWaveSync(); // leads to validation error - SharedMemoryDataRace-RaceOnAtomic
uint index = waveBinsIndex + WaveGetLaneIndex() % 16;
InterlockedAdd(sharedBins[index], 1);
//...I strongly simplify the code sample to focus on the issue, then it looses its initial purpose.
Expected behavior
The Validation Layer should recognize that GroupMemoryBarrierWithWaveSync establishes a valid happens-before relation between the non-atomic write (sharedBins[waveBinsIndex + WaveGetLaneIndex()] = 0) and the subsequent atomic operation (InterlockedAdd) for all invocations within that subgroup. No SharedMemoryDataRace-RaceOnAtomic warning/error should be emitted.
Valid Usage ID
Validation Error: [ SharedMemoryDataRace-RaceOnAtomic ] | MessageID = 0xe8f0e3c7
(Warning - This VUID has now been reported 10 times, which is the duplicate_message_limit value, this will be the last time reporting it).
[ Debug region: SplatsSortPass::DigitBinSizePerWGCountPass ] vkCmdDispatch(): A data race was detected on the shared memory variable "sharedBins" in local invocation index 131 while performing a atomic operation. (Likely against local invocation index 131)
The other access in this race was at:
Shader validation error occurred at /home/chtimy/Repositories/coconut/app/gaussian_splats/shaders/radix_sort/bins_count.slang:46:11
46: sharedBins[waveBinsIndex + WaveGetLaneIndex()] = 0;
^
Stage = Compute. Global invocation ID (x, y, z) = (131, 0, 0)
Command buffer (0)
Compute Dispatch Index 4
Shader Module (app/radix_sort/bins_count.slang:main)(0x26d000000026d) (internal ID 4)
Shader validation error occurred at /home/chtimy/Repositories/coconut/app/gaussian_splats/shaders/radix_sort/bins_count.slang:51:7
51: InterlockedAdd(sharedBins[waveBinsIndex + WaveGetLaneIndex()], 1);
^
Objects: 3
[0] VkQueue 0x55e251afbb10
[1] VkCommandBuffer 0x55e257204d20
[2] VkPipeline 0x2700000000270
Additional context
N/A
Environment:
Describe the Issue
The Vulkan Validation Layers trigger a false positive
SharedMemoryDataRace-RaceOnAtomicerror when non-atomic initialization of agroupsharedarray variable is followed by a subgroup-scoped control barrier (Slang function :GroupMemoryBarrierWithWaveSync) before subsequentInterlockedAddoperations within the same subgroup.Even though the execution and memory visibility are properly synchronized at the
Subgroupscope prior to atomic execution, the validation layer appears to require a full Workgroup-scoped execution barrier to clear the data race tracking state for Workgroup memory accesses.Here the minimal code :
I strongly simplify the code sample to focus on the issue, then it looses its initial purpose.
Expected behavior
The Validation Layer should recognize that
GroupMemoryBarrierWithWaveSyncestablishes a valid happens-before relation between the non-atomic write (sharedBins[waveBinsIndex + WaveGetLaneIndex()] = 0) and the subsequent atomic operation (InterlockedAdd) for all invocations within that subgroup. No SharedMemoryDataRace-RaceOnAtomic warning/error should be emitted.Valid Usage ID
Additional context
N/A