From 54b68eb343e9b29af024f4a5c4ac57e4e225c8fc Mon Sep 17 00:00:00 2001 From: Rostyslav Geyyer <46627076+geyyer@users.noreply.github.com> Date: Wed, 14 Jun 2023 16:06:56 -0500 Subject: [PATCH] Add generic kernel instances for ck::tensor_operation::device::DeviceGemmMultipleD (#741) * Add generic instance gemm_add_add_fastgelu * Add a client example for generic gemm_add_add_fastgelu * Update CMakeLists * Format * Format * Add generic instance gemm_add_fastgelu * Format * Add a gemm_add_fastgelu client example * Format * Add generic instance gemm_fastgelu * Format * Fix argument order * Add gemm_fastgelu client example * Add exceptions if argument is not supported --- .../02_gemm_add_add_fastgelu/CMakeLists.txt | 14 ++ .../gemm_add_add_fastgelu_generic.cpp | 176 ++++++++++++++++++ .../gemm_add_fastgelu_generic.cpp | 169 +++++++++++++++++ .../gemm_fastgelu_generic.cpp | 162 ++++++++++++++++ ...16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp | 14 ++ ...16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp | 14 ++ ...16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp | 14 ++ ...16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp | 14 ++ ...e_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp | 14 ++ ...e_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp | 14 ++ ...e_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp | 14 ++ ...e_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp | 14 ++ ..._shuffle_f16_f16_f16_km_kn_mn_instance.cpp | 12 ++ ..._shuffle_f16_f16_f16_km_nk_mn_instance.cpp | 12 ++ ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp | 12 ++ ..._shuffle_f16_f16_f16_mk_nk_mn_instance.cpp | 12 ++ 16 files changed, 681 insertions(+) create mode 100644 client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp create mode 100644 client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp create mode 100644 client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp diff --git a/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt index b7b724ccc4..ba29520222 100644 --- a/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt +++ b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt @@ -11,3 +11,17 @@ target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_ope add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu client_gemm_fastgelu) + +add_custom_target(client_gemm_fastgelu_generic_examples) + +add_executable(client_gemm_add_add_fastgelu_generic gemm_add_add_fastgelu_generic.cpp) +target_link_libraries(client_gemm_add_add_fastgelu_generic PRIVATE composable_kernel::device_operations) + +add_executable(client_gemm_add_fastgelu_generic gemm_add_fastgelu_generic.cpp) +target_link_libraries(client_gemm_add_fastgelu_generic PRIVATE composable_kernel::device_operations) + +add_executable(client_gemm_fastgelu_generic gemm_fastgelu_generic.cpp) +target_link_libraries(client_gemm_fastgelu_generic PRIVATE composable_kernel::device_operations) + +add_dependencies(client_gemm_fastgelu_generic_examples client_gemm_add_add_fastgelu_generic + client_gemm_add_fastgelu_generic client_gemm_fastgelu_generic) diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp new file mode 100644 index 0000000000..2ed942f0ad --- /dev/null +++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp @@ -0,0 +1,176 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp" + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu; + +using AElementOp = PassThrough; +using BElementOp = PassThrough; +using CDEElementOp = AddAddFastGelu; + +using ADataType = F16; +using BDataType = F16; +using D0DataType = F16; +using D1DataType = F16; +using EDataType = F16; + +using ALayout = Row; +using BLayout = Col; +using D0Layout = Row; +using D1Layout = Row; +using ELayout = Row; + +struct SimpleDeviceMem +{ + SimpleDeviceMem() = delete; + + SimpleDeviceMem(std::size_t mem_size) : p_mem_{} + { + (void)hipMalloc(static_cast(&p_mem_), mem_size); + } + + void* GetDeviceBuffer() { return p_mem_; } + + ~SimpleDeviceMem() { (void)hipFree(p_mem_); } + + void* p_mem_; +}; + +int main(int argc, char* argv[]) +{ + // GEMM shape + ck::index_t M = 3840; + ck::index_t N = 4096; + ck::index_t K = 4096; + + ck::index_t StrideA = 4096; + ck::index_t StrideB = 4096; + ck::index_t StrideD0 = 0; + ck::index_t StrideD1 = 4096; + ck::index_t StrideE = 4096; + + if(argc == 1) + { + // use default case + } + else if(argc == 9) + { + M = std::stoi(argv[1]); + N = std::stoi(argv[2]); + K = std::stoi(argv[3]); + + StrideA = std::stoi(argv[4]); + StrideB = std::stoi(argv[5]); + StrideD0 = std::stoi(argv[6]); + StrideD1 = std::stoi(argv[7]); + StrideE = std::stoi(argv[8]); + } + else + { + printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n"); + exit(0); + } + + auto f_matrix_space_size = + [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) { + using Layout = decltype(layout); + + if constexpr(std::is_same::value) + { + return (nRow - 1) * stride + nCol; + } + else + { + return (nCol - 1) * stride + nRow; + } + }; + + SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{})); + SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{})); + SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) * + f_matrix_space_size(M, N, StrideD0, D0Layout{})); + SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) * + f_matrix_space_size(M, N, StrideD1, D1Layout{})); + SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{})); + + using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD< + ALayout, + BLayout, + ck::Tuple, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::AddAddFastGelu>; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "found " << op_ptrs.size() << " instances" << std::endl; + + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto cde_element_op = CDEElementOp{}; + + // get generic instance + auto& op_ptr = op_ptrs[0]; + + std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString() + << std::endl; + + // run the generic instance + auto argument_ptr = + op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(), + b_device_buf.GetDeviceBuffer(), + std::array{d0_m_n_device_buf.GetDeviceBuffer(), + d1_m_n_device_buf.GetDeviceBuffer()}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + std::array{StrideD0, StrideD1}, + StrideE, + a_element_op, + b_element_op, + cde_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); + } + else + { + throw std::runtime_error( + "Generic instance should be suitable for various input lengths/strides"); + } + + std::cout << "Done" << std::endl; + + return 0; +} diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp new file mode 100644 index 0000000000..644b428fc9 --- /dev/null +++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp @@ -0,0 +1,169 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp" + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu; + +using AElementOp = PassThrough; +using BElementOp = PassThrough; +using CDEElementOp = AddFastGelu; + +using ADataType = F16; +using BDataType = F16; +using D0DataType = F16; +using EDataType = F16; + +using ALayout = Row; +using BLayout = Col; +using D0Layout = Row; +using ELayout = Row; + +struct SimpleDeviceMem +{ + SimpleDeviceMem() = delete; + + SimpleDeviceMem(std::size_t mem_size) : p_mem_{} + { + (void)hipMalloc(static_cast(&p_mem_), mem_size); + } + + void* GetDeviceBuffer() { return p_mem_; } + + ~SimpleDeviceMem() { (void)hipFree(p_mem_); } + + void* p_mem_; +}; + +int main(int argc, char* argv[]) +{ + // GEMM shape + ck::index_t M = 3840; + ck::index_t N = 4096; + ck::index_t K = 4096; + + ck::index_t StrideA = 4096; + ck::index_t StrideB = 4096; + ck::index_t StrideD0 = 0; + ck::index_t StrideE = 4096; + + if(argc == 1) + { + // use default case + } + else if(argc == 8) + { + M = std::stoi(argv[1]); + N = std::stoi(argv[2]); + K = std::stoi(argv[3]); + + StrideA = std::stoi(argv[4]); + StrideB = std::stoi(argv[5]); + StrideD0 = std::stoi(argv[6]); + StrideE = std::stoi(argv[7]); + } + else + { + printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, StrideE\n"); + exit(0); + } + + auto f_matrix_space_size = + [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) { + using Layout = decltype(layout); + + if constexpr(std::is_same::value) + { + return (nRow - 1) * stride + nCol; + } + else + { + return (nCol - 1) * stride + nRow; + } + }; + + SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{})); + SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{})); + SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) * + f_matrix_space_size(M, N, StrideD0, D0Layout{})); + SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{})); + + using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD< + ALayout, + BLayout, + ck::Tuple, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::AddFastGelu>; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "found " << op_ptrs.size() << " instances" << std::endl; + + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto cde_element_op = CDEElementOp{}; + + // get generic instance + auto& op_ptr = op_ptrs[0]; + + std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString() + << std::endl; + + // run the generic instance + auto argument_ptr = + op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(), + b_device_buf.GetDeviceBuffer(), + std::array{d0_m_n_device_buf.GetDeviceBuffer()}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + std::array{StrideD0}, + StrideE, + a_element_op, + b_element_op, + cde_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); + } + else + { + throw std::runtime_error( + "Generic instance should be suitable for various input lengths/strides"); + } + + std::cout << "Done" << std::endl; + + return 0; +} diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp new file mode 100644 index 0000000000..482e93b421 --- /dev/null +++ b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp @@ -0,0 +1,162 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp" + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using FastGelu = ck::tensor_operation::element_wise::FastGelu; + +using AElementOp = PassThrough; +using BElementOp = PassThrough; +using CDEElementOp = FastGelu; + +using ADataType = F16; +using BDataType = F16; +using EDataType = F16; + +using ALayout = Row; +using BLayout = Col; +using ELayout = Row; + +struct SimpleDeviceMem +{ + SimpleDeviceMem() = delete; + + SimpleDeviceMem(std::size_t mem_size) : p_mem_{} + { + (void)hipMalloc(static_cast(&p_mem_), mem_size); + } + + void* GetDeviceBuffer() { return p_mem_; } + + ~SimpleDeviceMem() { (void)hipFree(p_mem_); } + + void* p_mem_; +}; + +int main(int argc, char* argv[]) +{ + // GEMM shape + ck::index_t M = 3840; + ck::index_t N = 4096; + ck::index_t K = 4096; + + ck::index_t StrideA = 4096; + ck::index_t StrideB = 4096; + ck::index_t StrideE = 4096; + + if(argc == 1) + { + // use default case + } + else if(argc == 7) + { + M = std::stoi(argv[1]); + N = std::stoi(argv[2]); + K = std::stoi(argv[3]); + + StrideA = std::stoi(argv[4]); + StrideB = std::stoi(argv[5]); + StrideE = std::stoi(argv[6]); + } + else + { + printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideE\n"); + exit(0); + } + + auto f_matrix_space_size = + [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) { + using Layout = decltype(layout); + + if constexpr(std::is_same::value) + { + return (nRow - 1) * stride + nCol; + } + else + { + return (nCol - 1) * stride + nRow; + } + }; + + SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{})); + SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{})); + SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{})); + + using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD< + ALayout, + BLayout, + ck::Tuple<>, + ELayout, + ADataType, + BDataType, + ck::Tuple<>, + EDataType, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::FastGelu>; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "found " << op_ptrs.size() << " instances" << std::endl; + + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto cde_element_op = CDEElementOp{}; + + // get generic instance + auto& op_ptr = op_ptrs[0]; + + std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString() + << std::endl; + + // run the generic instance + auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(), + b_device_buf.GetDeviceBuffer(), + {}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + {}, + StrideE, + a_element_op, + b_element_op, + cde_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); + } + else + { + throw std::runtime_error( + "Generic instance should be suitable for various input lengths/strides"); + } + + std::cout << "Done" << std::endl; + + return 0; +} diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp index 7b65f8737a..125fbc21a2 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp @@ -36,6 +36,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[k, m], b[k, n], d0[m, n], d1[m, n] +using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 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, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances = std::tuple< // clang-format off @@ -139,6 +150,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn PassThrough, AddAddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp index a9fef5c60d..cc33692d73 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp @@ -36,6 +36,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[k, m], b[n, k], d0[m, n], d1[m, n] +using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 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, 16, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances = std::tuple< // clang-format off @@ -139,6 +150,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn PassThrough, AddAddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp index c3b1dfccab..704787a080 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp @@ -36,6 +36,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[m, k], b[k, n], d0[m, n], d1[m, n] +using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 2, 32, 32, 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, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances = std::tuple< // clang-format off @@ -139,6 +150,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn PassThrough, AddAddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp index 8338b34a4d..d64c9ec5e1 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp @@ -36,6 +36,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[m, k], b[n, k], d0[m, n], d1[m ,n] +using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Row_Tuple, Row, F16, F16, F32, F32, F16_F16_Tuple, F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 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, 16, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances = std::tuple< // clang-format off @@ -130,6 +141,9 @@ void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn PassThrough, AddAddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp index 357dc91aa1..e68bd8e7e4 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp @@ -21,6 +21,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0) // outout: e[m, n] // input: a[k, m], b[k, n], d0[m, n] +using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -123,6 +134,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_inst PassThrough, AddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp index 65b94d0879..5aaa2e8fe5 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp @@ -21,6 +21,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[k, m], b[n, k], d0[m, n], d1[m, n] +using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -123,6 +134,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_inst PassThrough, AddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp index 00cd07bab8..7a2a3dbaf3 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp @@ -21,6 +21,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[m, k], b[k, n], d0[m, n], d1[m, n] +using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -123,6 +134,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_inst PassThrough, AddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp index 9955a206b6..fa33609978 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp @@ -21,6 +21,17 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b), d0, d1) // outout: e[m, n] // input: a[m, k], b[n, k], d0[m, n], d1[m ,n] +using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_generic_instance = + std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -114,6 +125,9 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_inst PassThrough, AddFastGelu>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp index c5aa59f911..803c44c7f5 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp @@ -21,6 +21,16 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b)) // outout: e[m, n] // input: a[k, m], b[k, n] +using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_generic_instance = std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -122,6 +132,8 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances( PassThrough, FastGelu>>>& instances) { + add_device_operation_instances( + instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp index e71b269b8d..9b9ef3db24 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp @@ -21,6 +21,16 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b)) // outout: e[m, n] // input: a[k, m], b[k, n] +using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_generic_instance = std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -122,6 +132,8 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances( PassThrough, FastGelu>>>& instances) { + add_device_operation_instances( + instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp index fdf63f8112..1a0b6c9d1c 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp @@ -21,6 +21,16 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b)) // outout: e[m, n] // input: a[m, k], b[k, n] +using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_generic_instance = std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -122,6 +132,8 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances( PassThrough, FastGelu>>>& instances) { + add_device_operation_instances( + instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{}); add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp index 4c98a9d5e9..18b1c0e993 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp @@ -21,6 +21,16 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial // e = elementwise((a * b)) // outout: e[m, n] // input: a[m, k], b[n, k] +using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instance = std::tuple< + // clang-format off + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1> + // clang-format on + >; using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple< // clang-format off //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| @@ -113,6 +123,8 @@ void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances( PassThrough, FastGelu>>>& instances) { + add_device_operation_instances( + instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instance{}); add_device_operation_instances( instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{}); add_device_operation_instances(