Update AMD buffer coherency (#3403)

* Update AMD buffer coherency [AICK-421]

* fixes

* fix

* fixes

* fixes

* Add backward compatilibity

* fix

* fixes

* fix

* fix

* fix

* Update grouped_convolution_backward_weight_kernel.hpp
This commit is contained in:
Bartłomiej Kocot
2025-12-18 10:16:22 +01:00
committed by GitHub
parent 15e81397a4
commit 700b2ec9c0
11 changed files with 268 additions and 98 deletions

View File

@@ -545,6 +545,13 @@ struct GroupedConvolutionBackwardWeightKernel
return false;
}
#if defined(__gfx11__)
if constexpr(EpiloguePipeline::MemoryOperation != ck_tile::memory_operation_enum::set)
{
return false;
}
#endif
if constexpr(EpiloguePipeline_::MemoryOperation == memory_operation_enum::atomic_add)
{
if(kargs.k_batch == 1)
@@ -971,6 +978,12 @@ struct GroupedConvolutionBackwardWeightKernel
CK_TILE_DEVICE void operator()(GroupedConvBwdWeightKernelArgsSpecialized& kargs) const
{
#if defined(__gfx11__)
if constexpr(EpiloguePipeline::MemoryOperation != ck_tile::memory_operation_enum::set)
{
return;
}
#endif
if constexpr(GroupedConvTraitsType_::ExplicitGemm)
{
CallExplicitGemm(kargs);