ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 04:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Improve grouped conv kernel name generation (#3028)

Browse Source 
* Improve the grouped conv kernel name generation in CK Tile.

* Fix building CShuffle epilogue tests.

---------

Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>
This commit is contained in:
Ville Pietilä
2025-10-30 14:19:07 +02:00
committed by GitHub
parent 8c4cb4f9f4
commit 9ee9f4d2a3
5 changed files with 54 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
include/ck_tile/ops/common/utils.hpp
View File

@@ -19,6 +19,12 @@ template <> struct typeToStr<fp8_t> { static constexpr const char * name = "fp8"
template <> struct typeToStr<bf8_t> { static constexpr const char * name = "bf8"; };
template <> struct typeToStr<int8_t> { static constexpr const char * name = "int8"; };
template <> struct typeToStr<pk_int4_t> { static constexpr const char * name = "pk_int4"; };
template <memory_operation_enum MemOp> struct memOpToStr;
template <> struct memOpToStr<memory_operation_enum::set> { static constexpr const char * name = "set"; };
template <> struct memOpToStr<memory_operation_enum::atomic_add> { static constexpr const char * name = "atomic_add"; };
template <> struct memOpToStr<memory_operation_enum::atomic_max> { static constexpr const char * name = "atomic_max"; };
template <> struct memOpToStr<memory_operation_enum::add> { static constexpr const char * name = "add"; };
// clang-format on
template <typename ADataType_, typename BDataType_>
@@ -32,4 +38,10 @@ std::string gemm_prec_str()
return base_str;
}
template <memory_operation_enum MemOp_>
std::string mem_op_string()
{
return std::string(memOpToStr<MemOp_>::name);
}
} // namespace ck_tile

15
include/ck_tile/ops/epilogue/cshuffle_epilogue.hpp
View File

@@ -3,7 +3,9 @@
#pragma once
#include "ck_tile/host/concat.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common/utils.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/elementwise/unary_element_wise_operation.hpp"
@@ -123,6 +125,19 @@ struct CShuffleEpilogue
static_assert(NumDTensor == DsLayout::size(),
"The size of DsDataType and DsLayout should be the same");
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "CShuffleEpilogue",
concat('x', MWave, NWave),
concat('x', MPerXdl, NPerXdl, KPerXdl),
VectorSizeC,
isCTransposed ? "CTransposed" : "CNotTransposed",
mem_op_string<MemoryOperation>());
// clang-format on
}
/**
* @brief Get the vector store size for C tensor.
*

7
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_backward_data_kernel.hpp
View File

@@ -556,7 +556,12 @@ struct GroupedConvolutionBackwardDataKernel
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "grouped_convolution_backward_data", gemm_prec_str<InDataType, WeiDataType>, GemmPipeline::GetName());
return concat('_', "grouped_convolution_backward_data",
gemm_prec_str<InDataType, WeiDataType>(),
"gemm",
GemmPipeline::GetName(),
"epilogue",
EpiloguePipeline::GetName());
// clang-format on
}

19
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_backward_weight_kernel.hpp
View File

@@ -412,10 +412,21 @@ struct GroupedConvolutionBackwardWeightKernel
{
constexpr auto NumGroupsToMerge = GroupedConvTraitsType_::NumGroupsToMerge;
// clang-format off
if (NumGroupsToMerge > 1)
return concat('_', "grouped_convolution_backward_weight", gemm_prec_str<InDataType, WeiDataType>, GemmPipeline::GetName(), "merge", NumGroupsToMerge);
else
return concat('_', "grouped_convolution_backward_weight", gemm_prec_str<InDataType, WeiDataType>, GemmPipeline::GetName());
if (NumGroupsToMerge > 1) {
return concat('_', "grouped_convolution_backward_weight",
gemm_prec_str<InDataType, WeiDataType>(),
"gemm",
GemmPipeline::GetName(),
"epilogue",
EpiloguePipeline::GetName());
} else {
return concat('_', "grouped_convolution_backward_weight",
gemm_prec_str<InDataType, WeiDataType>(),
"gemm",
GemmPipeline::GetName(),
"epilogue",
EpiloguePipeline::GetName(), "merge", NumGroupsToMerge);
}
// clang-format on
}

7
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_forward_kernel.hpp
View File

@@ -452,7 +452,12 @@ struct GroupedConvolutionForwardKernel
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "grouped_convolution_forward", gemm_prec_str<InDataType, WeiDataType>, GemmPipeline::GetName());
return concat('_', "grouped_convolution_forward",
gemm_prec_str<InDataType, WeiDataType>(),
"gemm",
GemmPipeline::GetName(),
"epilogue",
EpiloguePipeline::GetName());
// clang-format on
}