mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
[CK TILE][AICK-439] Fix cshuffle epilogue wave per shuffle (#3364)
* [CK TILE] Fix cshufle epligoue wave per shuffle * Align shuffle per tile with smem * fixes * Fixes for double smem * fix
This commit is contained in:
@@ -1048,7 +1048,7 @@ struct GroupedConvolutionBackwardDataKernel
|
||||
|
||||
if constexpr(GemmPipeline::DoubleSmemBuffer == true)
|
||||
{
|
||||
__shared__ char smem_ptr_1[GetSmemSize()];
|
||||
__shared__ char smem_ptr_1[GemmPipeline::GetSmemSize()];
|
||||
if constexpr(!(EpiloguePipeline::MemoryOperation == memory_operation_enum::atomic_add &&
|
||||
GroupedConvTraitsType_::VectorSizeC % 2 != 0 &&
|
||||
is_any_of<OutDataType, fp16_t, bf16_t>::value))
|
||||
|
||||
@@ -1005,7 +1005,7 @@ struct GroupedConvolutionBackwardWeightKernel
|
||||
|
||||
if constexpr(GemmPipeline::DoubleSmemBuffer == true)
|
||||
{
|
||||
__shared__ char smem_ptr_1[GetSmemSize()];
|
||||
__shared__ char smem_ptr_1[GemmPipeline::GetSmemSize()];
|
||||
if constexpr(!(EpiloguePipeline::MemoryOperation ==
|
||||
memory_operation_enum::atomic_add &&
|
||||
GroupedConvTraitsType_::VectorSizeC % 2 != 0 &&
|
||||
|
||||
@@ -1184,7 +1184,7 @@ struct GroupedConvolutionForwardKernel
|
||||
|
||||
if constexpr(GemmPipeline::DoubleSmemBuffer == true)
|
||||
{
|
||||
__shared__ char smem_ptr_1[GetSmemSize()];
|
||||
__shared__ char smem_ptr_1[GemmPipeline::GetSmemSize()];
|
||||
if constexpr(!(EpiloguePipeline::MemoryOperation ==
|
||||
memory_operation_enum::atomic_add &&
|
||||
GroupedConvTraitsType_::VectorSizeC % 2 != 0 &&
|
||||
|
||||
Reference in New Issue
Block a user