ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Wmma support for gemm_bias_add_reduce (#3316)

Browse Source 
* Add tests for gemm_bias_add_reduce

* Initial working implementation

* Generalize implementation of reduce epilogue

* Add tests for all layouts

* Add instances

* Fix test archs

* Fix xdl bug

* Remove library/profiler duplications

* Fix num_byted error profiler

* Fix typos

* Fix copyright

[ROCm/composable_kernel commit: aad4cf0985]
This commit is contained in:
Enrico Degregori
2026-01-07 19:27:16 +01:00
committed by GitHub
parent d074af36c9
commit 6eab5bea54
15 changed files with 1424 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
View File

@@ -1,10 +1,15 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
# ONLY XDL_KERNELS
# ONLY XDL_AND_WMMA_KERNELS
add_instance_library(device_gemm_bias_add_reduce_instance
device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
)

85
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,85 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_wmma_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
// c[m, n] = a[k, m] * b[k, n]
using device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances =
std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| ELayout|AData| BData| EData|BiasData|D0Data| AccData| CShuffle| ReduceAcc| ReducePtrsGlobal| A| B| C| D0| ReduceOperations| Reduce| Reduce| Reduce| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| BlkGemm| BlkGemm|
//#####################################| | | | Type| Type| Type| Type| Type| Type| DataType| DataType| | Elementwise| Elementwise| Elementwise| Elementwise| | InElementwiseOperations| OutElementwiseOperations| GlobalMemory| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| PipeSched| PipelineVer|
//#####################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| | | | DataOperation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| | _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 2, 2, 16, 16, 8, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 8, 8, 16, 16, 8, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 2, 2, 16, 16, 2, 8, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 64, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 8, 8, 16, 16, 2, 8, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 2, 2, 16, 16, 4, 4, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 2, 2, 16, 16, 4, 2, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 2, 2, 16, 16, 4, 2, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 8, 8, 16, 16, 4, 2, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 2, 2, 16, 16, 4, 2, S< 8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8, S<16, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, S<16, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 64, 32, 2, 2, 16, 16, 2, 2, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 64, 32, 8, 8, 16, 16, 2, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 64, 128, 32, 2, 2, 16, 16, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 64, 128, 32, 8, 8, 16, 16, 2, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>
// // clang-format on
>;
void add_device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

84
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp Normal file
View File

@@ -0,0 +1,84 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_wmma_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
// c[m, n] = a[k, m] * b[n, k]
using device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances =
std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| ELayout|AData| BData| EData|BiasData|D0Data| AccData| CShuffle| ReduceAcc| ReducePtrsGlobal| A| B| C| D0| ReduceOperations| Reduce| Reduce| Reduce| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| BlkGemm| BlkGemm|
//#####################################| | | | Type| Type| Type| Type| Type| Type| DataType| DataType| | Elementwise| Elementwise| Elementwise| Elementwise| | InElementwiseOperations| OutElementwiseOperations| GlobalMemory| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| PipeSched| PipelineVer|
//#####################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| | | | DataOperation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| | _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 2, 8, 16, 16, 8, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 8, 8, 16, 16, 8, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 2, 8, 16, 16, 2, 8, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 8, 8, 16, 16, 2, 8, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 2, 8, 16, 16, 4, 4, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S< 4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 2, 8, 16, 16, 4, 2, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 2, 8, 16, 16, 4, 2, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 8, 8, 16, 16, 4, 2, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S< 4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 2, 8, 16, 16, 4, 2, S< 8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, S<16, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, S<16, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 64, 32, 2, 8, 16, 16, 2, 2, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 64, 32, 8, 8, 16, 16, 2, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 64, 128, 32, 2, 8, 16, 16, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 64, 128, 32, 8, 8, 16, 16, 2, 2, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S< 4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

84
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,84 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_wmma_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
// c[m, n] = a[m, k] * b[n, k]
using device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances =
std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| ELayout|AData| BData| EData|BiasData|D0Data| AccData| CShuffle| ReduceAcc| ReducePtrsGlobal| A| B| C| D0| ReduceOperations| Reduce| Reduce| Reduce| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| BlkGemm| BlkGemm|
//#####################################| | | | Type| Type| Type| Type| Type| Type| DataType| DataType| | Elementwise| Elementwise| Elementwise| Elementwise| | InElementwiseOperations| OutElementwiseOperations| GlobalMemory| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| PipeSched| PipelineVer|
//#####################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| | | | DataOperation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| | _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 8, 2, 16, 16, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 8, 8, 16, 16, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 8, 2, 16, 16, 2, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 64, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 8, 8, 16, 16, 2, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 8, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 8, 2, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 8, 2, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 8, 2, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8, S<16, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, S<16, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 64, 32, 8, 2, 16, 16, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 64, 128, 32, 8, 2, 16, 16, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 64, 128, 32, 8, 8, 16, 16, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

81
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp Normal file
View File

@@ -0,0 +1,81 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_wmma_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
// c[m, n] = a[m, k] * b[n, k]
using device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances =
std::tuple<
// clang-format off
//#####################################| ALayout| BLayout| ELayout|AData| BData| EData|BiasData|D0Data| AccData| CShuffle| ReduceAcc| ReducePtrsGlobal| A| B| C| D0| ReduceOperations| Reduce| Reduce| Reduce| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| BlkGemm| BlkGemm|
//#####################################| | | | Type| Type| Type| Type| Type| Type| DataType| DataType| | Elementwise| Elementwise| Elementwise| Elementwise| | InElementwiseOperations| OutElementwiseOperations| GlobalMemory| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| PipeSched| PipelineVer|
//#####################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| | | | DataOperation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| | _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| | |
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 256, 128, 32, 8, 8, 16, 16, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<32, 8>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 64, 64, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, S<16, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 64, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 128, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 128, 32, 128, 32, 8, 8, 16, 16, 1, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, S<32, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 64, 64, 32, 32, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8, S<32, 2>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGemmBiasAddReduce_Wmma_CShuffleV3< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp, GemmDefault, 64, 32, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, S<16, 4>, 4, 1, Intrawave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_mean_squaremean_wmma_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck