diff --git a/Dockerfile b/Dockerfile index 710db05c06..bc47768060 100644 --- a/Dockerfile +++ b/Dockerfile @@ -12,20 +12,20 @@ RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins RUN chmod 1777 /tmp RUN apt-get update RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl -RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.6" ]; then \ +RUN if [ "$ROCMVERSION" != "5.6" ]; then \ wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \ sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"; \ elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "" ]; then \ sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \ apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \ - amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914; \ - elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "rc3" ]; then \ + amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914 && \ + amdgpu-install -y --usecase=rocm --no-dkms; \ + elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "rc3" ] || [ "$compiler_version" = "amd-stg-open" ]; then \ sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.6-20.04-1_all.deb" && \ apt update && apt-get install -y ./amdgpu-install-internal_5.6-20.04-1_all.deb && \ sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.6 rel-45 > /etc/apt/sources.list.d/rocm-build.list' && \ - amdgpu-repo --amdgpu-build=1602498; \ + amdgpu-repo --amdgpu-build=1602498 && amdgpu-install -y --usecase=rocm --no-dkms; \ fi -RUN amdgpu-install -y --usecase=rocm --no-dkms RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add - RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" @@ -105,7 +105,7 @@ ENV compiler_commit=$compiler_commit RUN sh -c "echo compiler version = '$compiler_version'" RUN sh -c "echo compiler commit = '$compiler_commit'" -RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \ +RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \ git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \ cd llvm-project && mkdir build && cd build && \ cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \ @@ -113,7 +113,7 @@ RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler else echo "using the release compiler"; \ fi -RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \ +RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \ git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \ cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \ cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \ diff --git a/Jenkinsfile b/Jenkinsfile index fbff349fc3..8cfc5f8048 100644 --- a/Jenkinsfile +++ b/Jenkinsfile @@ -695,8 +695,8 @@ pipeline { } agent{ label rocmnode("gfx908 || gfx90a") } environment{ - setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """ - execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ + setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" """ + execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ } steps{ Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local') 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.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp index 5965e9d1d6..8d2a8c234a 100644 --- a/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp +++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp @@ -76,7 +76,7 @@ int main(int argc, char* argv[]) StrideA = std::stoi(argv[4]); StrideB = std::stoi(argv[5]); StrideD0 = std::stoi(argv[6]); - StrideE = std::stoi(argv[8]); + StrideE = std::stoi(argv[7]); } else { 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.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp index 319fdb0b03..c02df018fd 100644 --- a/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp +++ b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp @@ -72,7 +72,7 @@ int main(int argc, char* argv[]) StrideA = std::stoi(argv[4]); StrideB = std::stoi(argv[5]); - StrideE = std::stoi(argv[8]); + StrideE = std::stoi(argv[6]); } else { 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/client_example/16_convnd_fwd/common.hpp b/client_example/16_convnd_fwd/common.hpp index a6bb5aa65b..449c9466e8 100644 --- a/client_example/16_convnd_fwd/common.hpp +++ b/client_example/16_convnd_fwd/common.hpp @@ -141,14 +141,10 @@ bool run_grouped_conv_fwd(std::array( - {N, Di, Hi, Wi, G, C}, {K, Z, Y, X, G, C}, {N, Do, Ho, Wo, G, K}) + {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K}) ? EXIT_SUCCESS : EXIT_FAILURE; } diff --git a/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp index 43c98f1e9b..7e8c98b603 100644 --- a/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp +++ b/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp @@ -11,7 +11,7 @@ using WeiDataType = float; using OutDataType = float; using InLayout = ck::tensor_layout::convolution::NDHWGC; -using WeiLayout = ck::tensor_layout::convolution::KZYXGC; +using WeiLayout = ck::tensor_layout::convolution::GKZYXC; using OutLayout = ck::tensor_layout::convolution::NDHWGK; static constexpr ck::index_t NumDimSpatial = 3; @@ -38,7 +38,7 @@ int main() InLayout, WeiLayout, OutLayout>( - {N, Di, Hi, Wi, G, C}, {K, Z, Y, X, G, C}, {N, Do, Ho, Wo, G, K}) + {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K}) ? EXIT_SUCCESS : EXIT_FAILURE; } diff --git a/example/02_gemm_bilinear/CMakeLists.txt b/example/02_gemm_bilinear/CMakeLists.txt index eecec2437c..7ea4564d1b 100644 --- a/example/02_gemm_bilinear/CMakeLists.txt +++ b/example/02_gemm_bilinear/CMakeLists.txt @@ -1,6 +1,17 @@ -if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102") +list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102) +list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list1 AND target EQUAL 0) add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp) -endif() -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") + set(target 1) + endif() +endforeach() + +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list2 AND target EQUAL 0) add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp) -endif() + set(target 1) + endif() +endforeach() diff --git a/example/03_gemm_bias_relu/CMakeLists.txt b/example/03_gemm_bias_relu/CMakeLists.txt index 8834a910f0..2f5cba924d 100644 --- a/example/03_gemm_bias_relu/CMakeLists.txt +++ b/example/03_gemm_bias_relu/CMakeLists.txt @@ -1,3 +1,8 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() diff --git a/example/04_gemm_add_add_fastgelu/CMakeLists.txt b/example/04_gemm_add_add_fastgelu/CMakeLists.txt index a706830b68..447fa9871b 100644 --- a/example/04_gemm_add_add_fastgelu/CMakeLists.txt +++ b/example/04_gemm_add_add_fastgelu/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(example_gemm_add_add_fastgelu_xdl) add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp) @@ -16,4 +19,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4) endif(USE_BITINT_EXTENSION_INT4) add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/example/09_convnd_fwd/CMakeLists.txt b/example/09_convnd_fwd/CMakeLists.txt index 1bcf2d148d..90104c1630 100644 --- a/example/09_convnd_fwd/CMakeLists.txt +++ b/example/09_convnd_fwd/CMakeLists.txt @@ -1,12 +1,17 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp) add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp) add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp) add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp) # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp) -endif() + set(target 1) + endif() +endforeach() + add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp) add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp) add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp) - diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt b/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt index de26462f69..9577b4569a 100644 --- a/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt +++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(example_convnd_fwd_reduce_xdl) add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp) add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp) @@ -12,4 +15,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp) add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4) endif(USE_BITINT_EXTENSION_INT4) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/example/14_gemm_quantization/CMakeLists.txt b/example/14_gemm_quantization/CMakeLists.txt index 584333e7bb..72bdff5ab4 100644 --- a/example/14_gemm_quantization/CMakeLists.txt +++ b/example/14_gemm_quantization/CMakeLists.txt @@ -2,7 +2,12 @@ add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp) # xdlops -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp) add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt b/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt index 0f1ca777c5..a42b427c61 100644 --- a/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt +++ b/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(example_gemm_reduce_xdl) add_custom_target(example_gemm_reduce_xdl_max) add_custom_target(example_gemm_reduce_xdl_mean_meansquare) @@ -39,4 +42,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp) add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4) endif() -endif() + set(target 1) + endif() +endforeach() diff --git a/example/17_convnd_bwd_data/CMakeLists.txt b/example/17_convnd_bwd_data/CMakeLists.txt index ed95946c42..8ab9f37bb3 100644 --- a/example/17_convnd_bwd_data/CMakeLists.txt +++ b/example/17_convnd_bwd_data/CMakeLists.txt @@ -1,6 +1,11 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp) target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility) -endif() + set(target 1) + endif() +endforeach() add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp) target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility) diff --git a/example/18_batched_gemm_reduce/CMakeLists.txt b/example/18_batched_gemm_reduce/CMakeLists.txt index 0c3648dbfd..94ed129dc0 100644 --- a/example/18_batched_gemm_reduce/CMakeLists.txt +++ b/example/18_batched_gemm_reduce/CMakeLists.txt @@ -1,4 +1,8 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp) -endif() - + set(target 1) + endif() +endforeach() diff --git a/example/20_grouped_conv_bwd_weight/CMakeLists.txt b/example/20_grouped_conv_bwd_weight/CMakeLists.txt index 0ee39ac84e..db170decce 100644 --- a/example/20_grouped_conv_bwd_weight/CMakeLists.txt +++ b/example/20_grouped_conv_bwd_weight/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(example_grouped_conv_bwd_weight) add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp) @@ -6,7 +9,9 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16 example_grouped_conv_bwd_weight_xdl_bf16) -endif() + set(target 1) + endif() +endforeach() add_custom_target(example_grouped_conv_bwd_weight_dl) diff --git a/example/21_gemm_layernorm/CMakeLists.txt b/example/21_gemm_layernorm/CMakeLists.txt index 7f974221b7..ff870e7c6b 100644 --- a/example/21_gemm_layernorm/CMakeLists.txt +++ b/example/21_gemm_layernorm/CMakeLists.txt @@ -1,6 +1,11 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp) add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp) add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp) add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp) -endif() + set(target 1) + endif() +endforeach() diff --git a/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt b/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt index 9780a64cc2..1c07538b09 100644 --- a/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt +++ b/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt @@ -1,4 +1,9 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942) +list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102) + +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list1 AND target EQUAL 0) add_custom_target(example_grouped_conv_fwd_multiple_d) add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp) @@ -17,8 +22,15 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M endif() # USE_BITINT_EXTENSION_INT4 add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp) add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16) -endif() -if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102") - add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp) - add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp) -endif() + set(target 1) + endif() +endforeach() + +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list2 AND target EQUAL 0) + add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp) + add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp) + set(target 1) + endif() +endforeach() diff --git a/example/31_batched_gemm_gemm/CMakeLists.txt b/example/31_batched_gemm_gemm/CMakeLists.txt index dd9aef94a0..f8d139275d 100644 --- a/example/31_batched_gemm_gemm/CMakeLists.txt +++ b/example/31_batched_gemm_gemm/CMakeLists.txt @@ -1,12 +1,23 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942) +list(APPEND gpu_list2 gfx908 gfx90a) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list1 AND target EQUAL 0) add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp) add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp) add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp) - if(NOT GPU_TARGETS MATCHES "gfx940") - add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp) - endif() if(USE_BITINT_EXTENSION_INT4) add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp) endif(USE_BITINT_EXTENSION_INT4) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() + +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list2 AND target EQUAL 0) + add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp) + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/example/35_splitK_gemm/CMakeLists.txt b/example/35_splitK_gemm/CMakeLists.txt index f5a6ccb241..57ac33fc9a 100644 --- a/example/35_splitK_gemm/CMakeLists.txt +++ b/example/35_splitK_gemm/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(example_splitK_gemm_xdl) add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp) add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp) @@ -15,4 +18,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp) add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4) endif() -endif() + set(target 1) + endif() +endforeach() diff --git a/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt b/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt index 472d59c77d..de48093ac2 100644 --- a/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt +++ b/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt @@ -1,8 +1,13 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(example_grouped_conv_bwd_data) add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp) add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp) add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16) add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/example/40_conv2d_fwd_quantization/CMakeLists.txt b/example/40_conv2d_fwd_quantization/CMakeLists.txt index c12ab7a340..b82a013b58 100644 --- a/example/40_conv2d_fwd_quantization/CMakeLists.txt +++ b/example/40_conv2d_fwd_quantization/CMakeLists.txt @@ -1,9 +1,14 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_perchannel_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp) -endif() + set(target 1) + endif() +endforeach() # Conv perlayer quantization add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp) diff --git a/example/41_grouped_conv_conv_fwd/CMakeLists.txt b/example/41_grouped_conv_conv_fwd/CMakeLists.txt index 14f5c284aa..0c9df707b3 100644 --- a/example/41_grouped_conv_conv_fwd/CMakeLists.txt +++ b/example/41_grouped_conv_conv_fwd/CMakeLists.txt @@ -1,11 +1,22 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942) +list(APPEND gpu_list2 gfx908 gfx90a) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list1 AND target EQUAL 0) add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp) - if(NOT GPU_TARGETS MATCHES "gfx940") - add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp) - endif() if(USE_BITINT_EXTENSION_INT4) add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp) endif(USE_BITINT_EXTENSION_INT4) -endif() + set(target 1) + endif() +endforeach() + +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list2 AND target EQUAL 0) + add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp) + set(target 1) + endif() +endforeach() diff --git a/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt b/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt index b60789ea37..14432f6e23 100644 --- a/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt +++ b/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt @@ -1,3 +1,8 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp) -endif() + set(target 1) + endif() +endforeach() diff --git a/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp index cfb42c6e17..a90a6340a4 100644 --- a/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp +++ b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp @@ -121,7 +121,8 @@ using DeviceOpInstance = 2, // CShuffleNXdlPerWavePerShuffle S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock - MaskingSpec>; // MaskingSpecialization + MaskingSpec, // MaskingSpecialization + 1>; // Ref Gemm0: fp16 in, fp32 out using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm + +#include "ck/stream_config.hpp" +#include "ck/host_utility/hip_check_error.hpp" + +static int getAvailableComputeUnitCount(const StreamConfig& stream_config) +{ + constexpr int MAX_MASK_DWORDS = 64; + + // assume at most 64*32 = 2048 CUs + uint32_t cuMask[MAX_MASK_DWORDS]; + + for(int i = 0; i < MAX_MASK_DWORDS; i++) + cuMask[i] = 0; + + auto countSetBits = [](uint32_t dword) { + int count = 0; + + while(dword != 0) + { + if(dword & 0x1) + count++; + + dword = dword >> 1; + }; + + return (count); + }; + + hip_check_error(hipExtStreamGetCUMask(stream_config.stream_id_, MAX_MASK_DWORDS, &cuMask[0])); + + int ret = 0; + + for(int i = 0; i < MAX_MASK_DWORDS; i++) + ret += countSetBits(cuMask[i]); + + return (ret); +}; diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp index d00e19b440..d0de101c49 100644 --- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp @@ -48,7 +48,7 @@ __global__ void const Block2ETileMap block_2_etile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, @@ -418,7 +418,8 @@ struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm +#include + +#include "ck/utility/common_header.hpp" +#include "ck/tensor_description/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp" +#include "ck/host_utility/device_prop.hpp" +#include "ck/host_utility/kernel_launch.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +/* + * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix + * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly + * strided batched, but we can easily extend to other layouts. The returned offset can be either \p + * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB + * limitations. + * + * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2 + * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid + * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link + * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for + * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the + * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch. + */ + +template +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) +#endif + kernel_gemm_dl_multiple_d( + const ABDataType* __restrict__ p_a_grid, + const ABDataType* __restrict__ p_b_grid, + DsPointer p_ds_grid, + EDataType* __restrict__ p_e_grid, + const index_t batch_count, + const AElementwiseOperation a_element_op, + const BElementwiseOperation b_element_op, + const CDEElementwiseOperation cde_element_op, + const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1, + const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1, + const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11, + const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11, + const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch, + const Block2CTileMap block_2_ctile_map) +{ +// TODO: Enable for gfx90a after complier fix +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \ + defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \ + defined(__gfx1102__)) + + const index_t num_blocks_per_batch = + __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); + const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); + + const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx))); + const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx))); + const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx))); + + const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx); + + __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + DsPointer p_ds_grid_grp; + + static constexpr index_t NumDTensor = DsGridDesc_M0_M10_M11_N0_N10_N11::Size(); + + static_for<0, NumDTensor, 1>{}( + [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; }); + + GridwiseGemm::Run(p_a_grid + a_batch_offset, + p_b_grid + b_batch_offset, + p_ds_grid_grp, + p_e_grid + e_batch_offset, + p_shared, + a_element_op, + b_element_op, + cde_element_op, + a_grid_desc_k0_m0_m1_k1, + b_grid_desc_k0_n0_n1_k1, + ds_grid_desc_m0_m10_m11_n0_n10_n11, + e_grid_desc_m0_m10_m11_n0_n10_n11, + block_2_ctile_map, + integral_constant{}, + integral_constant{}); +#else + ignore = p_a_grid; + ignore = p_b_grid; + ignore = p_ds_grid; + ignore = p_e_grid; + ignore = batch_count; + ignore = a_element_op; + ignore = b_element_op; + ignore = cde_element_op; + ignore = a_grid_desc_k0_m0_m1_k1; + ignore = b_grid_desc_k0_n0_n1_k1; + ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11; + ignore = e_grid_desc_m0_m10_m11_n0_n10_n11; + ignore = compute_ptr_offset_of_batch; + ignore = block_2_ctile_map; + +#endif +} + +template && + is_same_v, + bool> = false> +struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD + +{ + using DeviceOp = DeviceBatchedGemmMultipleD_Dl; + static constexpr index_t NumDTensor = DsDataType::Size(); + + static constexpr auto I0 = Number<0>{}; + static constexpr auto I1 = Number<1>{}; + static constexpr auto I2 = Number<2>{}; + static constexpr auto I3 = Number<3>{}; + static constexpr auto I4 = Number<4>{}; + static constexpr auto I5 = Number<5>{}; + + static constexpr auto K1Number = Number{}; + + static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA) + { + const index_t K0 = K / K1; + + const auto a_grid_desc_m_k = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA)); + } + }(); + + if constexpr(GemmSpec == GemmSpecialization::MNPadding) + { + const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; + + return transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_right_pad_transform(M, PadM)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + else + { + return transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + } + + static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB) + { + const index_t K0 = K / K1; + + const auto b_grid_desc_k_n = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB)); + } + }(); + + if constexpr(GemmSpec == GemmSpecialization::MNPadding) + { + const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; + + return transform_tensor_descriptor( + b_grid_desc_k_n, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_right_pad_transform(N, PadN)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + else + { + return transform_tensor_descriptor( + b_grid_desc_k_n, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_pass_through_transform(N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + } + + template + static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE) + { + const auto c_grid_desc_m_n = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE)); + } + }(); + + if constexpr(GemmSpec == GemmSpecialization::MNPadding) + { + const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; + const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; + + return transform_tensor_descriptor( + c_grid_desc_m_n, + make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else + { + + return transform_tensor_descriptor( + c_grid_desc_m_n, + make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + } + + static auto MakeDsGridDescriptor_M_N(const std::array& MRaws, + const std::array& NRaws, + const std::array& DsStride) + { + return generate_tuple( + [&](auto i) { + using DLayout = remove_cvref_t>; + + return DeviceOp::MakeEGridDescriptor_M_N(MRaws[i], NRaws[i], DsStride[i]); + }, + Number{}); + } + + using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1)); + using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1)); + using DsGridDesc_M_N = decltype(MakeDsGridDescriptor_M_N({}, {}, {})); + using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1)); + + struct ComputePtrOffsetOfStridedBatch + { + ComputePtrOffsetOfStridedBatch(index_t BatchStrideA, + index_t BatchStrideB, + std::array BatchStrideDs, + index_t BatchStrideE) + : BatchStrideA_(BatchStrideA), + BatchStrideB_(BatchStrideB), + BatchStrideDs_(BatchStrideDs), + BatchStrideE_(BatchStrideE) + { + } + + __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideA_); + } + + __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideB_); + } + + __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const + { + std::array ds_offset; + static_for<0, NumDTensor, 1>{}([&](auto i) { + ds_offset[i] = g_idx * static_cast(BatchStrideDs_[i]); + }); + return ds_offset; + } + + __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideE_); + } + + private: + index_t BatchStrideA_; + index_t BatchStrideB_; + std::array BatchStrideDs_; + index_t BatchStrideE_; + }; + + // GridwiseGemm + using GridwiseGemm = + GridwiseGemmDlMultipleD_km_kn_mn; + + using AGridDesc_K0_M0_M1_K1 = + decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{})); + using BGridDesc_K0_N0_N1_K1 = + decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{})); + using DsGridDesc_M0_M10_M11_N0_N10_N11 = + decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{})); + using EGridDesc_M0_M10_M11_N0_N10_N11 = + decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{})); + using DefaultBlock2CTileMap = + decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{})); + + // Argument + struct Argument : public BaseArgument + { + Argument(const void* p_a_grid, + const void* p_b_grid, + std::array p_ds_grid, + void* p_e_grid, + index_t M, + index_t N, + index_t K, + index_t Batch, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideE, + index_t BatchStrideA, + index_t BatchStrideB, + const std::array& BatchStrideDs, + index_t BatchStrideE, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) + : p_a_grid_{static_cast(p_a_grid)}, + p_b_grid_{static_cast(p_b_grid)}, + p_ds_grid_{}, + p_e_grid_{static_cast(p_e_grid)}, + K_(K), + Batch_(Batch), + a_grid_desc_k0_m0_m1_k1_{}, + b_grid_desc_k0_n0_n1_k1_{}, + e_grid_desc_m0_m10_m11_n0_n10_n11_{}, + compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE}, + block_2_ctile_map_{}, + a_element_op_{a_element_op}, + b_element_op_{b_element_op}, + cde_element_op_{cde_element_op} + { + a_grid_desc_k0_m_k1_ = + DeviceBatchedGemmMultipleD_Dl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA); + b_grid_desc_k0_n_k1_ = + DeviceBatchedGemmMultipleD_Dl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB); + static_for<0, NumDTensor, 1>{}([&](auto i) { + using DLayout = remove_cvref_t>; + using DDataType = remove_cvref_t>; + + // D pointer + p_ds_grid_(i) = static_cast(p_ds_grid[i]); + + // D desc + ds_grid_desc_m_n_(i) = + DeviceOp::MakeEGridDescriptor_M_N(M, N, StrideDs[i]); + }); + e_grid_desc_m_n_ = + DeviceBatchedGemmMultipleD_Dl::MakeEGridDescriptor_M_N(M, N, StrideE); + + if(GridwiseGemm::CheckValidity( + a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, e_grid_desc_m_n_)) + { + a_grid_desc_k0_m0_m1_k1_ = + GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_); + b_grid_desc_k0_n0_n1_k1_ = + GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_); + + ds_grid_desc_m0_m10_m11_n0_n10_n11_ = + GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_); + + e_grid_desc_m0_m10_m11_n0_n10_n11_ = + GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_); + + block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_); + } + } + + // private: + const ADataType* p_a_grid_; + const BDataType* p_b_grid_; + typename GridwiseGemm::DsGridPointer p_ds_grid_; + EDataType* p_e_grid_; + + index_t K_; + + // Batch + index_t Batch_; + + AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_; + BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_; + DsGridDesc_M_N ds_grid_desc_m_n_; + EGridDesc_M_N e_grid_desc_m_n_; + + AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_; + BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_; + DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_; + EGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_; + + // for calculating batch offset + ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_; + + DefaultBlock2CTileMap block_2_ctile_map_; + + // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being. + AElementwiseOperation a_element_op_; + BElementwiseOperation b_element_op_; + CDEElementwiseOperation cde_element_op_; + }; + + // Invoker + struct Invoker : public BaseInvoker + { + using Argument = DeviceBatchedGemmMultipleD_Dl::Argument; + + float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) + { + { + std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{" + << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", " + << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", " + << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl; + + std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{" + << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", " + << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", " + << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl; + + std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", " + << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl; + } + + if(!GridwiseGemm::CheckValidity( + arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_)) + { + throw std::runtime_error( + "wrong! GridwiseGemmDlMultipleD_km_kn_mn has invalid setting"); + } + + const index_t grid_size = + GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0), + arg.e_grid_desc_m_n_.GetLength(I1)) * + arg.Batch_; + + auto launch_kernel = [&](auto has_main_k_block_loop, + auto has_double_tail_k_block_loop) { + constexpr bool has_main_loop = has_main_k_block_loop.value; + constexpr bool has_double_loop = has_double_tail_k_block_loop.value; + + const auto kernel = + kernel_gemm_dl_multiple_d; + + return launch_and_time_kernel(stream_config, + kernel, + dim3(grid_size), + dim3(BlockSize), + 0, + arg.p_a_grid_, + arg.p_b_grid_, + arg.p_ds_grid_, + arg.p_e_grid_, + arg.Batch_, + arg.a_element_op_, + arg.b_element_op_, + arg.cde_element_op_, + arg.a_grid_desc_k0_m0_m1_k1_, + arg.b_grid_desc_k0_n0_n1_k1_, + arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_, + arg.e_grid_desc_m0_m10_m11_n0_n10_n11_, + arg.compute_ptr_offset_of_batch_, + arg.block_2_ctile_map_); + }; + + const auto K0 = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0); + const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0); + const bool has_double_tail_k_block_loop = + GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0); + + if(has_main_k_block_loop && has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else if(has_main_k_block_loop && !has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else if(!has_main_k_block_loop && has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + } + + // polymorphic + float Run(const BaseArgument* p_arg, + const StreamConfig& stream_config = StreamConfig{}) override + { + return Run(*dynamic_cast(p_arg), stream_config); + } + }; + + static constexpr bool IsValidCompilationParameter() + { + // TODO: properly implement this check + return true; + } + + static bool IsSupportedArgument(const Argument& arg) + { + // TODO: Enable for gfx90a after complier fix + if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" || + ck::get_device_name() == "gfx1030" || ck::get_device_name() == "gfx940" || + ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" || + ck::get_device_name() == "gfx1102") + { + bool pass = true; + pass = pass && arg.K_ % K1 == 0; + + pass = pass && GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_, + arg.b_grid_desc_k0_n_k1_, + arg.e_grid_desc_m_n_); + + return pass; + } + else + { + return false; + } + } + + // polymorphic + bool IsSupportedArgument(const BaseArgument* p_arg) override + { + return IsSupportedArgument(*dynamic_cast(p_arg)); + } + + static auto MakeArgument(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_e, + index_t M, + index_t N, + index_t K, + index_t Batch, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideE, + index_t BatchStrideA, + index_t BatchStrideB, + const std::array& BatchStrideDs, + index_t BatchStrideE, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) + { + return Argument{p_a, + p_b, + p_ds, + p_e, + M, + N, + K, + Batch, + StrideA, + StrideB, + StrideDs, + StrideE, + BatchStrideA, + BatchStrideB, + BatchStrideDs, + BatchStrideE, + a_element_op, + b_element_op, + cde_element_op}; + } + + static auto MakeInvoker() { return Invoker{}; } + + // polymorphic + std::unique_ptr + MakeArgumentPointer(const void* p_a, + const void* p_b, + const std::array& p_ds, + void* p_e, + index_t M, + index_t N, + index_t K, + index_t Batch, + index_t StrideA, + index_t StrideB, + const std::array& StrideDs, + index_t StrideE, + index_t BatchStrideA, + index_t BatchStrideB, + const std::array& BatchStrideDs, + index_t BatchStrideE, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) override + { + return std::make_unique(p_a, + p_b, + p_ds, + p_e, + M, + N, + K, + Batch, + StrideA, + StrideB, + StrideDs, + StrideE, + BatchStrideA, + BatchStrideB, + BatchStrideDs, + BatchStrideE, + a_element_op, + b_element_op, + cde_element_op); + } + + // polymorphic + std::unique_ptr MakeInvokerPointer() override + { + return std::make_unique(Invoker{}); + } + + // polymorphic + std::string GetTypeString() const override + { + auto str = std::stringstream(); + + // clang-format off + str << "DeviceBatchedGemmMultipleD_Dl" + << "<" + << BlockSize << ", " + << MPerBlock << ", " + << NPerBlock << ", " + << K0PerBlock << ", " + << K1 << ", " + << M1PerThread << ", " + << N1PerThread << ", " + << KPerThread + << ">"; + // clang-format on + + return str.str(); + } +}; + +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp index e39d8f0696..d259ebffae 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp @@ -69,7 +69,7 @@ __global__ void const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); @@ -806,7 +806,8 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle static bool IsSupportedArgument(const Argument& arg) { if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || - ck::get_device_name() == "gfx940")) + ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || + ck::get_device_name() == "gfx942")) { return false; } diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp index eff503f650..1d52416d4d 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp @@ -60,7 +60,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp index e54c013cf9..ca72bcdd23 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp @@ -68,7 +68,7 @@ __global__ void const C0MatrixMask c0_matrix_mask) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); @@ -197,7 +197,8 @@ template + int D0sTransferSrcScalarPerVector = 4, + LoopScheduler LoopSched = LoopScheduler::Default> struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle : public DeviceBatchedGemmSoftmaxGemmPermute; + MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle, + D0sTransferSrcScalarPerVector>; // Argument // FIXME: constness @@ -530,6 +532,11 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle using D0DataType = remove_cvref_t>; // D0 pointer p_d0s_grid_(i) = static_cast(p_acc0_biases[i]); + // for check + d0s_nl_ns_lengths_strides_[i].push_back( + acc0_biases_gs_ms_ns_lengths[i][NumDimG + NumDimM]); + d0s_nl_ns_lengths_strides_[i].push_back( + acc0_biases_gs_ms_ns_strides[i][NumDimG + NumDimM]); }); if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_, @@ -608,6 +615,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle std::vector b_nz_kz_strides_; std::vector b1_nz_kz_strides_; std::vector c_mz_gemm1nz_strides_; + std::array, NumD0Tensor> d0s_nl_ns_lengths_strides_; index_t batch_count_; ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_; @@ -716,7 +724,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle #endif if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || - ck::get_device_name() == "gfx940")) + ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || + ck::get_device_name() == "gfx942")) { return false; } @@ -772,6 +781,20 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle { return false; } + for(int i = 0; i < NumD0Tensor; i++) + { + if(arg.d0s_nl_ns_lengths_strides_[i][1] == 1 && + arg.d0s_nl_ns_lengths_strides_[i][0] % D0sTransferSrcScalarPerVector != 0) + { + std::cout << "first" << std::endl; + return false; + } + if(arg.d0s_nl_ns_lengths_strides_[i][1] != 1 && D0sTransferSrcScalarPerVector != 1) + { + std::cout << "second" << std::endl; + return false; + } + } return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp index 9310d0752a..2405640ca9 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp @@ -63,7 +63,7 @@ __global__ void const C0MatrixMask c0_matrix_mask) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); @@ -614,7 +614,8 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle static bool IsSupportedArgument(const Argument& arg) { if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || - ck::get_device_name() == "gfx940")) + ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || + ck::get_device_name() == "gfx942")) { return false; } diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp index 87c58f371a..b6a0567fee 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp @@ -53,7 +53,7 @@ __global__ void kernel_batched_gemm_xdlops_v2r3(const typename DeviceOp::Argument karg) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / karg.Batch); const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp index b6f38698cd..dd57c90898 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp @@ -53,7 +53,7 @@ __global__ void const Block2ETileMap block_2_etile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, @@ -583,7 +583,8 @@ struct DeviceContractionMultipleD_Xdl_CShuffle static bool IsSupportedArgument(const Argument& arg) { if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || - ck::get_device_name() == "gfx940")) + ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || + ck::get_device_name() == "gfx942")) { return false; } diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp index fd8c88da9a..bc6582835e 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp @@ -56,7 +56,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / num_batches); const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp index c5f90e40fe..02ef29e32d 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp @@ -13,6 +13,7 @@ #include "ck/tensor_description/tensor_descriptor_helper.hpp" #include "ck/host_utility/kernel_launch.hpp" +#include "ck/host_utility/stream_utility.hpp" namespace ck { namespace tensor_operation { @@ -171,10 +172,7 @@ struct DeviceElementwise2dImpl : public DeviceElementwise 0, ""); static_assert(NumDim_n > 0, ""); @@ -192,34 +190,10 @@ struct DeviceElementwise2dImpl : public DeviceElementwise(out_dev_buffers[I.value]); }, Number{}); - - in_grid_2d_desc_tuple_ = generate_tuple( - [&](auto I) { - return MakeDescriptor_MN(lengths, - inStridesArray[I.value], - gridSize_, - blockSize_, - num_threads_m_, - num_threads_n_); - }, - Number{}); - - out_grid_2d_desc_tuple_ = generate_tuple( - [&](auto I) { - return MakeDescriptor_MN(lengths, - outStridesArray[I.value], - gridSize_, - blockSize_, - num_threads_m_, - num_threads_n_); - }, - Number{}); } InDataTypePointerTuple in_dev_buffers_; OutDataTypePointerTuple out_dev_buffers_; - InGrid2dDescTuple in_grid_2d_desc_tuple_; - OutGrid2dDescTuple out_grid_2d_desc_tuple_; std::array lengths_; std::array, NumInput> inStridesArray_; @@ -227,15 +201,38 @@ struct DeviceElementwise2dImpl : public DeviceElementwise{}); + + auto out_grid_2d_desc_tuple = generate_tuple( + [&](auto I) { + return MakeDescriptor_MN(arg.lengths_, + arg.outStridesArray_[I.value], + gridSize, + arg.blockSize_, + num_threads_m, + num_threads_n); + }, + Number{}); + const auto kernel = kernel_elementwise_2d(out_dev_buffers[I.value]); }, Number{}); - - in_grid_1d_desc_tuple_ = generate_tuple( - [&](auto I) { - return MakeDescriptor_M( - lengths, inStridesArray[I.value], gridSize_, blockSize_); - }, - Number{}); - - out_grid_1d_desc_tuple_ = generate_tuple( - [&](auto I) { - return MakeDescriptor_M( - lengths, outStridesArray[I.value], gridSize_, blockSize_); - }, - Number{}); } InDataTypePointerTuple in_dev_buffers_; OutDataTypePointerTuple out_dev_buffers_; - InGrid1dDescTuple in_grid_1d_desc_tuple_; - OutGrid1dDescTuple out_grid_1d_desc_tuple_; std::array lengths_; std::array, NumInput> inStridesArray_; @@ -187,13 +171,28 @@ struct DeviceElementwiseImpl ElementwiseOperation elementwise_op_; index_t blockSize_; - index_t gridSize_; }; struct Invoker : public BaseInvoker { float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) { + index_t gridSize = getAvailableComputeUnitCount(stream_config); + + auto in_grid_1d_desc_tuple = generate_tuple( + [&](auto I) { + return MakeDescriptor_M( + arg.lengths_, arg.inStridesArray_[I.value], gridSize, arg.blockSize_); + }, + Number{}); + + auto out_grid_1d_desc_tuple = generate_tuple( + [&](auto I) { + return MakeDescriptor_M( + arg.lengths_, arg.outStridesArray_[I.value], gridSize, arg.blockSize_); + }, + Number{}); + const auto kernel = kernel_elementwise_1d( @@ -856,7 +856,8 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle static bool IsSupportedArgument(const Argument& arg) { if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || - ck::get_device_name() == "gfx940")) + ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || + ck::get_device_name() == "gfx942")) { return false; } diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp index be174e599e..3a0be837b7 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp @@ -61,7 +61,7 @@ __global__ void const Block2ETileMap block_2_etile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, @@ -556,7 +556,8 @@ struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle static bool IsSupportedArgument(const Argument& arg) { if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || - ck::get_device_name() == "gfx940")) + ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || + ck::get_device_name() == "gfx942")) { return false; } diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp index 9aaa68b0cb..873f1bc511 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp @@ -52,7 +52,7 @@ __global__ void const Block2ETileMap block_2_etile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, @@ -492,7 +492,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD || is_same_v || is_same_v || is_same_v)) diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp index 02458bf02a..a9294f0344 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp @@ -136,7 +136,7 @@ __global__ void const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) // offset base pointer for each work-group const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); @@ -685,7 +685,8 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle return false; } } - else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940") + else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940" || + get_device_name() == "gfx941" || get_device_name() == "gfx942") { if constexpr(!(is_same_v || is_same_v || is_same_v || is_same_v)) diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp index 22be58259a..0190b3cee6 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp @@ -41,7 +41,7 @@ __global__ void { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \ defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \ - defined(__gfx1102__) || defined(__gfx940__)) + defined(__gfx1102__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; const index_t block_id = get_block_1d_id(); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp index 3900047567..bc9b3a4dcc 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp @@ -39,7 +39,7 @@ __global__ void const CDEElementwiseOperation c_element_op) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; const index_t block_id = get_block_1d_id(); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp index 1ac9969f86..74f38b9db2 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp @@ -35,7 +35,7 @@ __global__ void const index_t group_count) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte(); __shared__ uint8_t p_shared[shared_size]; diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp index 74171ea9d9..135b9da6a0 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp @@ -80,7 +80,8 @@ template + int D0sTransferSrcScalarPerVector = 4, + PipelineVersion PipelineVer = PipelineVersion::v1> struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle { static_assert(LoopSched == LoopScheduler::Default, @@ -621,13 +622,13 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 = make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId I1, // NBlockID - I1, // MRepeat - I1, // NRepeat - I1, // MWaveId - I1, // NWaveId - I1, // MPerXdl - I1, // NGroupNum - I1, // NInputNum + m0, // MRepeat + n0, // NRepeat + m1, // MWaveId + n1, // NWaveId + m2, // MPerXdl + n2, // NGroupNum + n3, // NInputNum n4)); // registerNum auto d0s_thread_buf = generate_tuple( @@ -644,9 +645,6 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle const auto wave_id = GetGemm0WaveIdx(); const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63 - constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed( - make_tuple(Number{}, Number{}, n2, n4)); - auto d0s_threadwise_copy = generate_tuple( [&](auto i) { using D0DataType = remove_cvref_t>; @@ -655,10 +653,19 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle D0DataType, decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]), decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5), - Sequence, + Sequence, Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>, 9, - n4, + D0sTransferSrcScalarPerVector, 1, false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], make_multi_index(block_work_idx[I0], // MBlockId @@ -884,62 +891,35 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle // multiple d if constexpr(NumD0Tensor) { - static_for<0, MXdlPerWave, 1>{}([&](auto mr) { - static_for<0, NXdlPerWave, 1>{}([&](auto nr) { - static_for<0, n2, 1>{}([&](auto groupid) { - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).Run( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - d0s_grid_buf[i], - d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, - make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0), - d0s_thread_buf(i)); - }); + static_assert(NXdlPerWave == n0); + static_assert(MXdlPerWave == m0); - static_for<0, n4, 1>{}([&](auto i) { - constexpr index_t c_offset = acc0_thread_desc.CalculateOffset( - make_tuple(mr, nr, groupid, i)); + static_for<0, NumD0Tensor, 1>{}([&](auto i) { + d0s_threadwise_copy(i).Run(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], + d0s_grid_buf[i], + d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, + make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0), + d0s_thread_buf(i)); + }); + static_for<0, m0 * n0 * n2 * n4, 1>{}([&](auto i) { + // get reference to src data + const auto src_data_refs = generate_tie( + // return type should be lvalue + [&](auto iSrc) -> const auto& { return d0s_thread_buf[iSrc][i]; }, + Number{}); - // get reference to src data - const auto src_data_refs = generate_tie( - // return type should be lvalue - [&](auto iSrc) -> const auto& { - return d0s_thread_buf[iSrc][i]; - }, - Number{}); + // get reference to dst data + auto dst_data_refs = generate_tie( + // return type should be lvalue + [&](auto) -> auto& { return acc_thread_buf(i); }, + Number<2>{}); - // get reference to dst data - auto dst_data_refs = generate_tie( - // return type should be lvalue - [&](auto) -> auto& { - return acc_thread_buf(Number{}); - }, - Number<2>{}); - - unpack2(c0de_element_op, dst_data_refs, src_data_refs); - }); - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).MoveSrcSliceWindow( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0)); - }); - }); - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).MoveSrcSliceWindow( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0)); - }); - }); - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).MoveSrcSliceWindow( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 0, 1, -NXdlPerWave, 0, 0, 0, 0, 0, 0)); - }); + unpack2(c0de_element_op, dst_data_refs, src_data_refs); }); static_for<0, NumD0Tensor, 1>{}([&](auto i) { d0s_threadwise_copy(i).MoveSrcSliceWindow( d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 1, -MXdlPerWave, 0, 0, 0, 0, 0, 0, 0)); + make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0)); }); } else diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp index c5c09e909e..533559adf1 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp @@ -67,7 +67,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp index bb3e6a80bf..3d1fbb73bd 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp @@ -55,7 +55,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940)) + defined(__gfx940) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp index d5552656e3..7f1fb21d33 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp @@ -25,7 +25,7 @@ __global__ void kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run( @@ -46,7 +46,7 @@ __global__ void typename GridwiseGemm::Problem problem) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, p_b_grid, p_c_grid, p_shared, problem); diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp index d805c9fa2b..8675350094 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp @@ -58,7 +58,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; // TODO ANT: separate into MMA + Epilogue diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp index da7ad1cacf..0fc18bb923 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp @@ -166,7 +166,7 @@ __global__ void const CBlockClusterAdaptor c_block_cluster_adaptor) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp index e9881d645b..b12bcee0f4 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp @@ -45,7 +45,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp index bd18fdb10d..7b8bbd301e 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp @@ -36,7 +36,7 @@ __global__ void const CGridDesc_M_N c_grid_desc_m_n) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid, @@ -64,7 +64,7 @@ __global__ void kernel_gemm_xdlops_v2r3(const typename GridwiseGemm::Argument karg) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; const auto a_grid_desc_k0_m_k1 = diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp index 33a4f2b2c5..19fbee727f 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp @@ -43,7 +43,7 @@ __global__ void const CBlockClusterAdaptor c_block_cluster_adaptor) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) constexpr index_t shared_block_size = GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB); diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp index a4b320ddfa..4cd70cc902 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp @@ -31,7 +31,7 @@ __global__ void const Block2CTileMap& b2c_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte(); __shared__ uint8_t p_shared[shared_size]; diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp index df543c063f..d090ba54d8 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp @@ -47,7 +47,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run( diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp index ec98fc9c93..2c09e80fdc 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp @@ -50,7 +50,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run( diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp index 3a752dd749..c1bc6a8fec 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp @@ -54,7 +54,7 @@ __global__ void const Block2CTileMap block_2_ctile_map) { #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__)) + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run( diff --git a/include/ck/utility/amd_buffer_addressing.hpp b/include/ck/utility/amd_buffer_addressing.hpp index ef3d2032c5..38ee76d883 100644 --- a/include/ck/utility/amd_buffer_addressing.hpp +++ b/include/ck/utility/amd_buffer_addressing.hpp @@ -286,7 +286,22 @@ llvm_amdgcn_raw_buffer_atomic_max_fp64(double vdata, int soffset, // dst_wave_addr_offset int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64"); -template +// memory coherency bit for buffer store/load instruction +// check ISA manual for each GFX target +// e.g. for +// https://www.amd.com/system/files/TechDocs/instinct-mi200-cdna2-instruction-set-architecture.pdf, +// page 67~68 +enum struct AmdBufferCoherenceEnum +{ + DefaultCoherence = 0, // default value + GLC = 1, + SLC = 2, + GLC_SLC = 3, +}; + +template __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_wave_buffer_resource, index_t src_thread_addr_offset, index_t src_wave_addr_offset) @@ -305,28 +320,37 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w // use fp32 load to mimic fp64 load if constexpr(N == 1) { - const float2_t tmp = llvm_amdgcn_raw_buffer_load_fp32x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + const float2_t tmp = + llvm_amdgcn_raw_buffer_load_fp32x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); } else if constexpr(N == 2) { - const float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + const float4_t tmp = + llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); } else if constexpr(N == 4) { - const float4_t f32_0 = llvm_amdgcn_raw_buffer_load_fp32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + const float4_t f32_0 = + llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); const float4_t f32_1 = llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 4 * sizeof(float), - 0); + static_cast(coherence)); vector_type tmp; tmp.AsType()(Number<0>{}) = bit_cast(f32_0); @@ -339,31 +363,40 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w { if constexpr(N == 1) { - return llvm_amdgcn_raw_buffer_load_fp32( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_fp32(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 2) { - return llvm_amdgcn_raw_buffer_load_fp32x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_fp32x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 4) { - return llvm_amdgcn_raw_buffer_load_fp32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 8) { vector_type tmp; - tmp.AsType()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_fp32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + tmp.AsType()(Number<0>{}) = + llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); tmp.AsType()(Number<1>{}) = llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 4 * sizeof(float), - 0); + static_cast(coherence)); return tmp.AsType()(Number<0>{}); } @@ -372,24 +405,32 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w { if constexpr(N == 1) { - return llvm_amdgcn_raw_buffer_load_fp16( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_fp16(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 2) { - return llvm_amdgcn_raw_buffer_load_fp16x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_fp16x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 4) { - return llvm_amdgcn_raw_buffer_load_fp16x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_fp16x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 8) { // use fp32 load to mimic fp16 load - float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); } @@ -398,23 +439,31 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w { if constexpr(N == 1) { - return llvm_amdgcn_raw_buffer_load_i16( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i16(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 2) { - return llvm_amdgcn_raw_buffer_load_i16x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i16x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 4) { - return llvm_amdgcn_raw_buffer_load_i16x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i16x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 8) { - int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); } @@ -423,31 +472,40 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w { if constexpr(N == 1) { - return llvm_amdgcn_raw_buffer_load_i32( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i32(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 2) { - return llvm_amdgcn_raw_buffer_load_i32x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i32x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 4) { - return llvm_amdgcn_raw_buffer_load_i32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 8) { vector_type tmp; - tmp.AsType()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_i32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + tmp.AsType()(Number<0>{}) = + llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); tmp.AsType()(Number<1>{}) = llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 4 * sizeof(int32_t), - 0); + static_cast(coherence)); return tmp.AsType()(Number<0>{}); } } @@ -455,17 +513,23 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w { if constexpr(N == 1) { - return llvm_amdgcn_raw_buffer_load_i8( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i8(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); } else if constexpr(N == 2) { #if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE - return llvm_amdgcn_raw_buffer_load_i8x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i8x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); #else - int16_t tmp = llvm_amdgcn_raw_buffer_load_i16( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + int16_t tmp = llvm_amdgcn_raw_buffer_load_i16(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); #endif @@ -473,11 +537,15 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w else if constexpr(N == 4) { #if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE - return llvm_amdgcn_raw_buffer_load_i8x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + return llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); #else - int32_t tmp = llvm_amdgcn_raw_buffer_load_i32( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + int32_t tmp = llvm_amdgcn_raw_buffer_load_i32(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); #endif @@ -487,19 +555,24 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w #if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE vector_type tmp; - tmp.AsType()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_i8x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + tmp.AsType()(Number<0>{}) = + llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); tmp.AsType()(Number<1>{}) = llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 4 * sizeof(int8_t), - 0); + static_cast(coherence)); return tmp.AsType()(Number<0>{}); #else - int32x2_t tmp = llvm_amdgcn_raw_buffer_load_i32x2( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + int32x2_t tmp = llvm_amdgcn_raw_buffer_load_i32x2(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); #endif @@ -509,31 +582,36 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w #if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE vector_type tmp; - tmp.AsType()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_i8x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + tmp.AsType()(Number<0>{}) = + llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); tmp.AsType()(Number<1>{}) = llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 4 * sizeof(int8_t), - 0); + static_cast(coherence)); tmp.AsType()(Number<2>{}) = llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 8 * sizeof(int8_t), - 0); + static_cast(coherence)); tmp.AsType()(Number<3>{}) = llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset + 12 * sizeof(int8_t), - 0); + static_cast(coherence)); return tmp.AsType()(Number<0>{}); #else - int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4( - src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); + int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource, + src_thread_addr_offset, + src_wave_addr_offset, + static_cast(coherence)); return bit_cast(tmp); #endif @@ -541,7 +619,9 @@ __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_w } } -template +template __device__ void amd_buffer_store_impl(const typename vector_type::type src_thread_data, int32x4_t dst_wave_buffer_resource, index_t dst_thread_addr_offset, @@ -565,7 +645,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 2) { @@ -573,7 +653,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } } else if constexpr(is_same::value) @@ -584,7 +664,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 2) { @@ -592,7 +672,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 4) { @@ -600,7 +680,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } } else if constexpr(is_same::value) @@ -611,7 +691,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 2) { @@ -619,7 +699,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 4) { @@ -627,7 +707,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 8) { @@ -638,19 +718,19 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType()[Number<1>{}], dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset + 4 * sizeof(half_t), - 0); + static_cast(coherence)); #else llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast(src_thread_data), dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); #endif } } @@ -662,7 +742,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 2) { @@ -670,7 +750,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 4) { @@ -678,7 +758,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 8) { @@ -688,13 +768,13 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType()[Number<1>{}], dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset + 4 * sizeof(bhalf_t), - 0); + static_cast(coherence)); } } else if constexpr(is_same::value) @@ -705,7 +785,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 2) { @@ -713,7 +793,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 4) { @@ -721,7 +801,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } } else if constexpr(is_same::value) @@ -732,7 +812,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 2) { @@ -741,13 +821,13 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); #else llvm_amdgcn_raw_buffer_store_i16(bit_cast(src_thread_data), dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); #endif } else if constexpr(N == 4) @@ -757,13 +837,13 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); #else llvm_amdgcn_raw_buffer_store_i32(bit_cast(src_thread_data), dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); #endif } else if constexpr(N == 8) @@ -772,7 +852,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } else if constexpr(N == 16) { @@ -780,7 +860,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type::type src dst_wave_buffer_resource, dst_thread_addr_offset, dst_wave_addr_offset, - 0); + static_cast(coherence)); } } } @@ -1012,7 +1092,9 @@ __device__ void amd_buffer_atomic_max_impl(const typename vector_type::typ // 1) p_src_wave must point to global memory space // 2) p_src_wave must be a wavewise pointer. // It is user's responsibility to make sure that is true. -template +template __device__ typename vector_type_maker::type::type amd_buffer_load_invalid_element_return_zero(const T* p_src_wave, index_t src_thread_element_offset, @@ -1032,10 +1114,10 @@ amd_buffer_load_invalid_element_return_zero(const T* p_src_wave, #if CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK uint32_t src_addr_shift = src_thread_element_valid ? 0 : 0x80000000; - return amd_buffer_load_impl( + return amd_buffer_load_impl( src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0); #else - vector_t tmp = amd_buffer_load_impl( + vector_t tmp = amd_buffer_load_impl( src_wave_buffer_resource, src_thread_addr_offset, 0); return src_thread_element_valid ? tmp : vector_t(0); @@ -1046,7 +1128,9 @@ amd_buffer_load_invalid_element_return_zero(const T* p_src_wave, // 1) p_src_wave must point to global memory space // 2) p_src_wave must be a wavewise pointer. // It is user's responsibility to make sure that is true. -template +template __device__ typename vector_type_maker::type::type amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave, index_t src_thread_element_offset, @@ -1064,7 +1148,7 @@ amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave, constexpr index_t vector_size = scalar_type::vector_size; - vector_t tmp = amd_buffer_load_impl( + vector_t tmp = amd_buffer_load_impl( src_wave_buffer_resource, src_thread_addr_offset, 0); return src_thread_element_valid ? tmp : vector_t(customized_value); @@ -1074,7 +1158,9 @@ amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave, // 1) p_dst_wave must point to global memory // 2) p_dst_wave must be a wavewise pointer. // It is user's responsibility to make sure that is true. -template +template __device__ void amd_buffer_store(const typename vector_type_maker::type::type src_thread_data, T* p_dst_wave, const index_t dst_thread_element_offset, @@ -1093,12 +1179,12 @@ __device__ void amd_buffer_store(const typename vector_type_maker::type::t #if CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x80000000; - amd_buffer_store_impl( + amd_buffer_store_impl( src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0); #else if(dst_thread_element_valid) { - amd_buffer_store_impl( + amd_buffer_store_impl( src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0); } #endif diff --git a/include/ck/utility/amd_wave_read_first_lane.hpp b/include/ck/utility/amd_wave_read_first_lane.hpp index 4652ce7a74..741b2975af 100644 --- a/include/ck/utility/amd_wave_read_first_lane.hpp +++ b/include/ck/utility/amd_wave_read_first_lane.hpp @@ -7,6 +7,7 @@ #include "ck/utility/functional2.hpp" #include "ck/utility/math.hpp" +#include #include #include #include @@ -14,29 +15,83 @@ namespace ck { namespace detail { -template -struct get_unsigned_int; +template +struct get_carrier; template <> -struct get_unsigned_int<1> +struct get_carrier<1> { using type = uint8_t; }; template <> -struct get_unsigned_int<2> +struct get_carrier<2> { using type = uint16_t; }; template <> -struct get_unsigned_int<4> +struct get_carrier<3> +{ + using type = class carrier + { + using value_type = uint32_t; + + std::array bytes; + static_assert(sizeof(bytes) <= sizeof(value_type)); + + // replacement of host std::copy_n() + template + __device__ static OutputIterator copy_n(InputIterator from, Size size, OutputIterator to) + { + if(0 < size) + { + *to = *from; + ++to; + for(Size count = 1; count < size; ++count) + { + *to = *++from; + ++to; + } + } + + return to; + } + + // method to trigger template substitution failure + __device__ carrier(const carrier& other) noexcept + { + copy_n(other.bytes.begin(), bytes.size(), bytes.begin()); + } + + public: + __device__ carrier& operator=(value_type value) noexcept + { + copy_n(reinterpret_cast(&value), bytes.size(), bytes.begin()); + + return *this; + } + + __device__ operator value_type() const noexcept + { + std::byte result[sizeof(value_type)]; + + copy_n(bytes.begin(), bytes.size(), result); + + return *reinterpret_cast(result); + } + }; +}; +static_assert(sizeof(get_carrier<3>::type) == 3); + +template <> +struct get_carrier<4> { using type = uint32_t; }; -template -using get_unsigned_int_t = typename get_unsigned_int::type; +template +using get_carrier_t = typename get_carrier::type; } // namespace detail @@ -61,7 +116,7 @@ __device__ auto amd_wave_read_first_lane(const Object& obj) constexpr Size CompleteSgprCopyBoundary = ObjectSize - RemainedSize; for(Size offset = 0; offset < CompleteSgprCopyBoundary; offset += SgprSize) { - using Sgpr = detail::get_unsigned_int_t; + using Sgpr = detail::get_carrier_t; *reinterpret_cast(to_obj + offset) = amd_wave_read_first_lane(*reinterpret_cast(from_obj + offset)); @@ -69,9 +124,9 @@ __device__ auto amd_wave_read_first_lane(const Object& obj) if constexpr(0 < RemainedSize) { - using Carrier = detail::get_unsigned_int_t; + using Carrier = detail::get_carrier_t; - *reinterpret_cast(to_obj + CompleteSgprCopyBoundary) = amd_wave_read_first_lane( + *reinterpret_cast(to_obj + CompleteSgprCopyBoundary) = amd_wave_read_first_lane( *reinterpret_cast(from_obj + CompleteSgprCopyBoundary)); } diff --git a/include/ck/utility/amd_xdlops.hpp b/include/ck/utility/amd_xdlops.hpp index f4a04e2814..d00b7cd078 100644 --- a/include/ck/utility/amd_xdlops.hpp +++ b/include/ck/utility/amd_xdlops.hpp @@ -344,7 +344,7 @@ struct intrin_mfma_f64_16x16x4f64<16, 16> template __device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c) { -#if defined(__gfx90a__) || defined(__gfx940__) +#if defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) reg_c.template AsType()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64( reg_a, reg_b, reg_c.template AsType()[Number<0>{}], 0, 0, 0); #else diff --git a/include/ck/utility/dynamic_buffer.hpp b/include/ck/utility/dynamic_buffer.hpp index 9ea0d6c001..02d61f34ed 100644 --- a/include/ck/utility/dynamic_buffer.hpp +++ b/include/ck/utility/dynamic_buffer.hpp @@ -19,7 +19,8 @@ namespace ck { template + bool InvalidElementUseNumericalZeroValue, + AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence> struct DynamicBuffer { using type = T; @@ -77,13 +78,16 @@ struct DynamicBuffer if constexpr(InvalidElementUseNumericalZeroValue) { - return amd_buffer_load_invalid_element_return_zero, t_per_x>( + return amd_buffer_load_invalid_element_return_zero, + t_per_x, + coherence>( p_data_, i, is_valid_element, element_space_size_); } else { return amd_buffer_load_invalid_element_return_customized_value, - t_per_x>( + t_per_x, + coherence>( p_data_, i, is_valid_element, element_space_size_, invalid_element_value_); } } @@ -173,7 +177,7 @@ struct DynamicBuffer { constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector; - amd_buffer_store, t_per_x>( + amd_buffer_store, t_per_x, coherence>( x, p_data_, i, is_valid_element, element_space_size_); } else if constexpr(GetAddressSpace() == AddressSpaceEnum::Lds && @@ -376,14 +380,19 @@ struct DynamicBuffer __host__ __device__ static constexpr bool IsDynamicBuffer() { return true; } }; -template +template __host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size) { - return DynamicBuffer{p, element_space_size}; + return DynamicBuffer{ + p, element_space_size}; } template < AddressSpaceEnum BufferAddressSpace, + AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence, typename T, typename ElementSpaceSize, typename X, @@ -391,7 +400,7 @@ template < __host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value) { - return DynamicBuffer{ + return DynamicBuffer{ p, element_space_size, invalid_element_value}; } diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp new file mode 100644 index 0000000000..ae12f4c7ac --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp @@ -0,0 +1,337 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances); + +template +struct DeviceOperationInstanceFactory> +{ + using DeviceOp = DeviceBatchedGemmMultiD; + + static auto GetInstances() + { + std::vector> op_ptrs; + + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances( + op_ptrs); + } + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances( + op_ptrs); + } + } + + return op_ptrs; + } +}; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp index a82ec543c6..627a5ae2aa 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp @@ -245,11 +245,11 @@ void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances( PassThrough, PassThrough>>>& instances); -// grouped conv3d forward, NDHWGC/KZYXGC/NDHWGK -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances( +// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances( std::vector && - is_same_v && is_same_v) + is_same_v && is_same_v) { if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances(op_ptrs); + add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs); } else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances(op_ptrs); + add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs); } else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances(op_ptrs); + add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs); } else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances(op_ptrs); + add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(op_ptrs); } } diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt new file mode 100644 index 0000000000..fda55a9303 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt @@ -0,0 +1,18 @@ +add_instance_library(device_batched_gemm_multi_d_instance + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp +) diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..3fe9f78b2a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..4ab22bb03a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,84 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..80c890cdb3 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..647c583036 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..3ce582f1f9 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..34c29d0380 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..e0da8e3d5a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..3edd3f78c6 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.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 Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..33234edc17 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 4, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 4, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 4, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 4, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 4, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 4, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 4, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 4, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 4, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..16107e1bde --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 4, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 4, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 4, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 4, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..3e4bdb0172 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..8107717162 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..0e943c88c3 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..ea5e7c562d --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..000a4b0130 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..24fb67619a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.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/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp index a541a2d224..498bf58fb3 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp @@ -41,10 +41,11 @@ template , ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, @@ -58,8 +59,9 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_ DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp index 5e481603ca..744bd6456d 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp @@ -41,10 +41,11 @@ template , ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, #if CK_WORKAROUND_SWDEV_388832 @@ -60,6 +61,7 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16 DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp index 501ea85f60..b342612d1c 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp @@ -45,6 +45,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_ // #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | // #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | // #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, @@ -58,8 +59,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_ DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp index 91ab541bf9..3fd0c07370 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp @@ -45,6 +45,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_ // #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | // #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | // #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, #if CK_WORKAROUND_SWDEV_388832 @@ -60,8 +61,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_ DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on >; 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( diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt index 90efc09ee7..cd209dbf9e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt @@ -4,8 +4,8 @@ add_instance_library(device_grouped_conv3d_fwd_instance device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp - device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instance.cpp - device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instance.cpp - device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instance.cpp - device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instance.cpp + device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp + device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp + device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp + device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp ) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp similarity index 91% rename from library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instance.cpp rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp index 5e5dbc53cd..d2cacf4539 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -26,7 +26,7 @@ template using S = ck::Sequence; using NDHWGC = ck::tensor_layout::convolution::NDHWGC; -using KZYXGC = ck::tensor_layout::convolution::KZYXGC; +using GKZYXC = ck::tensor_layout::convolution::GKZYXC; using NDHWGK = ck::tensor_layout::convolution::NDHWGK; using PassThrough = ck::tensor_operation::element_wise::PassThrough; @@ -43,7 +43,7 @@ static constexpr auto ConvFwd1x1S1P0 = static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; // in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k] -using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances = +using device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances = std::tuple< // clang-format off // Default @@ -51,64 +51,64 @@ using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances = //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, // Filter1x1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, // Filter1x1Stride1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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> // clang-format on >; -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances( std::vector>>& instances) { add_device_operation_instances( - instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances{}); + instances, device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp similarity index 91% rename from library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instance.cpp rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp index a88fe4af63..6354f01112 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp @@ -26,7 +26,7 @@ template using S = ck::Sequence; using NDHWGC = ck::tensor_layout::convolution::NDHWGC; -using KZYXGC = ck::tensor_layout::convolution::KZYXGC; +using GKZYXC = ck::tensor_layout::convolution::GKZYXC; using NDHWGK = ck::tensor_layout::convolution::NDHWGK; using PassThrough = ck::tensor_operation::element_wise::PassThrough; @@ -43,7 +43,7 @@ static constexpr auto ConvFwd1x1S1P0 = static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; // in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k] -using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances = +using device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances = std::tuple< // clang-format off // Default @@ -51,64 +51,64 @@ using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances = //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, // Filter1x1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, // Filter1x1Stride1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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> // clang-format on >; -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( std::vector>>& instances) { add_device_operation_instances( - instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances{}); + instances, device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp similarity index 91% rename from library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instance.cpp rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp index f6e7e5b28e..381ac5ef08 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp @@ -25,7 +25,7 @@ template using S = ck::Sequence; using NDHWGC = ck::tensor_layout::convolution::NDHWGC; -using KZYXGC = ck::tensor_layout::convolution::KZYXGC; +using GKZYXC = ck::tensor_layout::convolution::GKZYXC; using NDHWGK = ck::tensor_layout::convolution::NDHWGK; using PassThrough = ck::tensor_operation::element_wise::PassThrough; @@ -42,7 +42,7 @@ static constexpr auto ConvFwd1x1S1P0 = static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; // in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k] -using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances = +using device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances = std::tuple< // clang-format off // Default @@ -50,64 +50,64 @@ using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances = //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, // Filter1x1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, // Filter1x1Stride1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4> // clang-format on >; -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances( std::vector>>& instances) { add_device_operation_instances( - instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances{}); + instances, device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp similarity index 91% rename from library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instance.cpp rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp index 3d303a3fa7..6bd53d8693 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp @@ -23,7 +23,7 @@ template using S = ck::Sequence; using NDHWGC = ck::tensor_layout::convolution::NDHWGC; -using KZYXGC = ck::tensor_layout::convolution::KZYXGC; +using GKZYXC = ck::tensor_layout::convolution::GKZYXC; using NDHWGK = ck::tensor_layout::convolution::NDHWGK; using PassThrough = ck::tensor_operation::element_wise::PassThrough; @@ -40,71 +40,71 @@ static constexpr auto ConvFwd1x1S1P0 = static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; // in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k] -using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances = std::tuple< +using device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances = std::tuple< // clang-format off // Default //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, // Filter1x1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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>, // Filter1x1Stride1Pad0 //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, 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>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 3, NDHWGC, GKZYXC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 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> // clang-format on >; -void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances( +void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances( std::vector>>& instances) { add_device_operation_instances( - instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances{}); + instances, device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances{}); } } // namespace instance diff --git a/profiler/README.md b/profiler/README.md index 400a64a396..7a1fb29113 100644 --- a/profiler/README.md +++ b/profiler/README.md @@ -76,3 +76,30 @@ e_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1} .... Best Perf: 211.405 ms, 41.6077 TFlops, 15.2372 GB/s ``` + +## Profile batched gemm multiple D kernels +```bash +#arg1: tensor operation (batched_gemm_multi_d=Batched GEMM multi D); +#arg2: data type (0: fp16; 1: int8) +#arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n]; +# 1: A[g, m, k] * B[g, n, k] = C[g, m, n]; +# 2: A[g, k, m] * B[g, k, n] = C[g, m, n]; +# 3: A[g, k, m] * B[g, n, k] = C[g, m, n]) +#arg4: verification (0: no; 1: yes) +#arg5: initialization (0: no init; 1: integer value; 2: decimal value) +#arg6: print tensor value (0: no; 1: yes) +#arg7: time kernel (0=n0, 1=yes) +#arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount + +################ op datatype layout verify init log time M N K StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount +./bin/ckProfiler batched_gemm_multi_d 0 1 0 0 0 1 4096 4096 4096 4096 4096 4096 16777216 16777216 16777216 16 +``` + +Result (Radeon RX 6800 XT) +```bash +arg.a_grid_desc_k0_m0_m1_k1_{2048, 4096, 2} +arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2} +arg.e_grid_desc_m_n_{ 4096, 4096} +.... +Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s +``` diff --git a/profiler/include/profiler/profile_batched_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_impl.hpp index cdc94aa9a6..936c22f5d8 100644 --- a/profiler/include/profiler/profile_batched_gemm_impl.hpp +++ b/profiler/include/profiler/profile_batched_gemm_impl.hpp @@ -8,9 +8,11 @@ #include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp" #include "ck/library/utility/check_err.hpp" #include "ck/library/utility/device_memory.hpp" @@ -27,7 +29,11 @@ template + typename CLayout, + typename AElementOp, + typename BElementOp, + typename CElementOp, + typename DeviceOp> bool profile_batched_gemm_impl(int do_verification, int init_method, bool do_log, @@ -88,10 +94,6 @@ bool profile_batched_gemm_impl(int do_verification, b_g_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); } - using AElementOp = ck::tensor_operation::element_wise::PassThrough; - using BElementOp = ck::tensor_operation::element_wise::PassThrough; - using CElementOp = ck::tensor_operation::element_wise::PassThrough; - const auto a_element_op = AElementOp{}; const auto b_element_op = BElementOp{}; const auto c_element_op = CElementOp{}; @@ -124,16 +126,6 @@ bool profile_batched_gemm_impl(int do_verification, b_device_buf.ToDevice(b_g_k_n.mData.data()); c_device_buf.ToDevice(c_g_m_n_device_result.mData.data()); - using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm; - // get device op instances const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< DeviceOp>::GetInstances(); @@ -148,23 +140,62 @@ bool profile_batched_gemm_impl(int do_verification, // profile device op instances for(auto& op_ptr : op_ptrs) { - auto argument_ptr = - op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), - static_cast(b_device_buf.GetDeviceBuffer()), - static_cast(c_device_buf.GetDeviceBuffer()), - M, - N, - K, - StrideA, - StrideB, - StrideC, - BatchStrideA, - BatchStrideB, - BatchStrideC, - BatchCount, - ck::tensor_operation::element_wise::PassThrough{}, - ck::tensor_operation::element_wise::PassThrough{}, - ck::tensor_operation::element_wise::PassThrough{}); + std::unique_ptr argument_ptr; + // false branch for multi d dl kernel + if constexpr(std::is_same< + DeviceOp, + ck::tensor_operation::device::DeviceBatchedGemm>::value) + { + + argument_ptr = + op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + StrideA, + StrideB, + StrideC, + BatchStrideA, + BatchStrideB, + BatchStrideC, + BatchCount, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}); + } + else + { + argument_ptr = + op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + {}, + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + BatchCount, + StrideA, + StrideB, + {}, + StrideC, + BatchStrideA, + BatchStrideB, + {}, + BatchStrideC, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}); + } auto invoker_ptr = op_ptr->MakeInvokerPointer(); diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index c9fccc2588..6f768e0ae1 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -34,6 +34,7 @@ set(PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp profile_contraction_bilinear.cpp profile_contraction_scale.cpp + profile_batched_gemm_multi_d.cpp ) set(PROFILER_EXECUTABLE ckProfiler) @@ -77,5 +78,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgel target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance) - +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) diff --git a/profiler/src/profile_batched_gemm.cpp b/profiler/src/profile_batched_gemm.cpp index dc83e25b48..222532b7bb 100644 --- a/profiler/src/profile_batched_gemm.cpp +++ b/profiler/src/profile_batched_gemm.cpp @@ -10,6 +10,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" #include "profiler_operation_registry.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + enum struct GemmMatrixLayout { MK_KN_MN, // 0 @@ -78,55 +80,72 @@ int profile_batched_gemm(int argc, char* argv[]) using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; - auto profile = [&](auto a_type, - auto b_type, - auto c_type, - auto a_layout, - auto b_layout, - auto c_layout) { - using ADataType = decltype(a_type); - using BDataType = decltype(b_type); - using CDataType = decltype(c_type); + auto profile = + [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) { + using ADataType = decltype(a_type); + using BDataType = decltype(b_type); + using CDataType = decltype(c_type); - using ALayout = decltype(a_layout); - using BLayout = decltype(b_layout); - using CLayout = decltype(c_layout); + using ALayout = decltype(a_layout); + using BLayout = decltype(b_layout); + using CLayout = decltype(c_layout); - const int DefaultStrideA = ck::is_same_v ? K : M; - const int DefaultStrideB = ck::is_same_v ? N : K; - const int DefaultStrideC = ck::is_same_v ? N : M; + const int DefaultStrideA = ck::is_same_v ? K : M; + const int DefaultStrideB = ck::is_same_v ? N : K; + const int DefaultStrideC = ck::is_same_v ? N : M; - const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA; - const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; - const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; + const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA; + const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; + const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; - const int DefaultBatchStrideA = (ck::is_same_v ? M : K) * StrideA_; - const int DefaultBatchStrideB = (ck::is_same_v ? K : N) * StrideB_; - const int DefaultBatchStrideC = (ck::is_same_v ? M : N) * StrideC_; + const int DefaultBatchStrideA = (ck::is_same_v ? M : K) * StrideA_; + const int DefaultBatchStrideB = (ck::is_same_v ? K : N) * StrideB_; + const int DefaultBatchStrideC = (ck::is_same_v ? M : N) * StrideC_; - const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA; - const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB; - const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC; + const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA; + const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB; + const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC; - bool pass = ck::profiler:: - profile_batched_gemm_impl( - do_verification, - init_method, - do_log, - time_kernel, - M, - N, - K, - BatchStrideA_, - BatchStrideB_, - BatchStrideC_, - StrideA_, - StrideB_, - StrideC_, - BatchCount); + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; - return pass ? 0 : 1; - }; + using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm; + + bool pass = ck::profiler::profile_batched_gemm_impl(do_verification, + init_method, + do_log, + time_kernel, + M, + N, + K, + BatchStrideA_, + BatchStrideB_, + BatchStrideC_, + StrideA_, + StrideB_, + StrideC_, + BatchCount); + + return pass ? 0 : 1; + }; if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN) { diff --git a/profiler/src/profile_batched_gemm_multi_d.cpp b/profiler/src/profile_batched_gemm_multi_d.cpp new file mode 100644 index 0000000000..98b462d950 --- /dev/null +++ b/profiler/src/profile_batched_gemm_multi_d.cpp @@ -0,0 +1,190 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include +#include + +#include "profiler/profile_batched_gemm_impl.hpp" +#include "profiler_operation_registry.hpp" + +#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp" + +enum struct GemmMatrixLayout +{ + MK_KN_MN, // 0 + MK_NK_MN, // 1 + KM_KN_MN, // 2 + KM_NK_MN, // 3 +}; + +enum struct GemmDataType +{ + F16_F16_F16, // 0 + INT8_INT8_INT8, // 1 +}; + +#define OP_NAME "batched_gemm_multi_d" +#define OP_DESC "Batched GEMM multi D" + +int profile_batched_gemm_multi_d(int argc, char* argv[]) +{ + if(argc != 18) + { + // clang-format off + printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); + printf("arg2: data type (0: fp16; 1: int8)\n"); + printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n"); + printf(" 1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n"); + printf(" 2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n"); + printf(" 3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n"); + printf("arg4: verification (0: no; 1: yes)\n"); + printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"); + printf("arg6: print tensor value (0: no; 1: yes)\n"); + printf("arg7: time kernel (0=n0, 1=yes)\n"); + printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n"); + // clang-format on + exit(1); + } + + const auto data_type = static_cast(std::stoi(argv[2])); + const auto layout = static_cast(std::stoi(argv[3])); + const bool do_verification = std::stoi(argv[4]); + const int init_method = std::stoi(argv[5]); + const bool do_log = std::stoi(argv[6]); + const bool time_kernel = std::stoi(argv[7]); + + const int M = std::stoi(argv[8]); + const int N = std::stoi(argv[9]); + const int K = std::stoi(argv[10]); + + const int StrideA = std::stoi(argv[11]); + const int StrideB = std::stoi(argv[12]); + const int StrideC = std::stoi(argv[13]); + + const int BatchStrideA = std::stoi(argv[14]); + const int BatchStrideB = std::stoi(argv[15]); + const int BatchStrideC = std::stoi(argv[16]); + + const int BatchCount = std::stoi(argv[17]); + + using F16 = ck::half_t; + using INT8 = int8_t; + + using Row = ck::tensor_layout::gemm::RowMajor; + using Col = ck::tensor_layout::gemm::ColumnMajor; + + auto profile = + [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) { + using ADataType = decltype(a_type); + using BDataType = decltype(b_type); + using CDataType = decltype(c_type); + using DsDataType = ck::Tuple<>; + + using ALayout = decltype(a_layout); + using BLayout = decltype(b_layout); + using CLayout = decltype(c_layout); + using DsLayout = ck::Tuple<>; + + const int DefaultStrideA = ck::is_same_v ? K : M; + const int DefaultStrideB = ck::is_same_v ? N : K; + const int DefaultStrideC = ck::is_same_v ? N : M; + + const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA; + const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; + const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; + + const int DefaultBatchStrideA = (ck::is_same_v ? M : K) * StrideA_; + const int DefaultBatchStrideB = (ck::is_same_v ? K : N) * StrideB_; + const int DefaultBatchStrideC = (ck::is_same_v ? M : N) * StrideC_; + + const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA; + const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB; + const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC; + + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; + + using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmMultiD; + + bool pass = ck::profiler::profile_batched_gemm_impl(do_verification, + init_method, + do_log, + time_kernel, + M, + N, + K, + BatchStrideA_, + BatchStrideB_, + BatchStrideC_, + StrideA_, + StrideB_, + StrideC_, + BatchCount); + + return pass ? 0 : 1; + }; + + if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN) + { + return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN) + { + return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN) + { + return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN) + { + return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{}); + } + else + { + std::cout << "this data_type & layout is not implemented" << std::endl; + + return 1; + } +} + +REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_multi_d); diff --git a/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh index 426f68d443..33352a6220 100755 --- a/script/cmake-ck-dev.sh +++ b/script/cmake-ck-dev.sh @@ -12,7 +12,7 @@ cmake -save-temps=$PWD" \ -D CMAKE_BUILD_TYPE=Release \ -D BUILD_DEV=ON \ --D GPU_TARGETS="gfx908;gfx90a;gfx940" \ +-D GPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" \ -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \ -D USE_BITINT_EXTENSION_INT4=OFF \ ${MY_PROJECT_SOURCE} diff --git a/script/cmake-ck-release.sh b/script/cmake-ck-release.sh index 787eabbf96..03d422478e 100755 --- a/script/cmake-ck-release.sh +++ b/script/cmake-ck-release.sh @@ -11,7 +11,7 @@ cmake -D CMAKE_CXX_FLAGS="-O3" \ -D CMAKE_BUILD_TYPE=Release \ -D BUILD_DEV=OFF \ --D GPU_TARGETS="gfx908;gfx90a;gfx940" \ +-D GPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" \ -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \ -D USE_BITINT_EXTENSION_INT4=OFF \ ${MY_PROJECT_SOURCE} diff --git a/test/batched_gemm/CMakeLists.txt b/test/batched_gemm/CMakeLists.txt index 5552a6c98e..12142c33e7 100644 --- a/test/batched_gemm/CMakeLists.txt +++ b/test/batched_gemm/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp) target_link_libraries(test_batched_gemm_fp16 PRIVATE utility) target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance) @@ -14,4 +17,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp) target_link_libraries(test_batched_gemm_int8 PRIVATE utility) target_link_libraries(test_batched_gemm_int8 PRIVATE device_batched_gemm_instance) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/test/batched_gemm/batched_gemm_bf16.cpp b/test/batched_gemm/batched_gemm_bf16.cpp index fa1652e997..5d12a1e956 100644 --- a/test/batched_gemm/batched_gemm_bf16.cpp +++ b/test/batched_gemm/batched_gemm_bf16.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = ck::bhalf_t; using BDataType = ck::bhalf_t; @@ -12,6 +14,8 @@ using CDataType = ck::bhalf_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM bf16: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm/batched_gemm_fp16.cpp b/test/batched_gemm/batched_gemm_fp16.cpp index 3df4912aa0..a2b61d951a 100644 --- a/test/batched_gemm/batched_gemm_fp16.cpp +++ b/test/batched_gemm/batched_gemm_fp16.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = ck::half_t; using BDataType = ck::half_t; @@ -12,6 +14,8 @@ using CDataType = ck::half_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM fp16: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm/batched_gemm_fp32.cpp b/test/batched_gemm/batched_gemm_fp32.cpp index 2d808441b9..2b18d166e6 100644 --- a/test/batched_gemm/batched_gemm_fp32.cpp +++ b/test/batched_gemm/batched_gemm_fp32.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = float; using BDataType = float; @@ -12,6 +14,8 @@ using CDataType = float; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM fp32: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm/batched_gemm_int8.cpp b/test/batched_gemm/batched_gemm_int8.cpp index ed233a5ada..f607eaa84b 100644 --- a/test/batched_gemm/batched_gemm_int8.cpp +++ b/test/batched_gemm/batched_gemm_int8.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = int8_t; using BDataType = int8_t; @@ -12,6 +14,8 @@ using CDataType = int8_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM int8: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm_gemm/CMakeLists.txt b/test/batched_gemm_gemm/CMakeLists.txt index eff3391750..858efd8379 100644 --- a/test/batched_gemm_gemm/CMakeLists.txt +++ b/test/batched_gemm_gemm/CMakeLists.txt @@ -1,7 +1,11 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(test_batched_gemm_gemm) - add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp) target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance) add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/test/batched_gemm_multi_d/CMakeLists.txt b/test/batched_gemm_multi_d/CMakeLists.txt new file mode 100644 index 0000000000..45a306551f --- /dev/null +++ b/test/batched_gemm_multi_d/CMakeLists.txt @@ -0,0 +1,5 @@ +# TODO: Enable for gfx90a after complier fix +if(NOT GPU_TARGETS MATCHES "gfx90a") + add_gtest_executable(test_batched_gemm_multi_d test_batched_gemm_multi_d.cpp) + target_link_libraries(test_batched_gemm_multi_d PRIVATE utility device_batched_gemm_multi_d_instance) +endif() diff --git a/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp b/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp new file mode 100644 index 0000000000..4a82654034 --- /dev/null +++ b/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp @@ -0,0 +1,74 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include + +#include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp" + +namespace { +using F16 = ck::half_t; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using Empty_Tuple = ck::Tuple<>; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +template +class TestBatchedGemmMultiD : public ::testing::Test +{ + protected: + using ALayout = std::tuple_element_t<0, Tuple>; + using BLayout = std::tuple_element_t<1, Tuple>; + using CLayout = std::tuple_element_t<2, Tuple>; + + static constexpr int M = 512; + static constexpr int N = 256; + static constexpr int K = 128; + static constexpr int BatchCount = 3; + + template + void Run() + { + using namespace ck::tensor_operation::device; + + const bool pass = + ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + EXPECT_TRUE(pass); + } +}; + +using KernelTypes = ::testing::Types, + std::tuple, + std::tuple, + std::tuple>; +} // namespace + +TYPED_TEST_SUITE(TestBatchedGemmMultiD, KernelTypes); + +TYPED_TEST(TestBatchedGemmMultiD, f16) { this->template Run(); } + +TYPED_TEST(TestBatchedGemmMultiD, int8) { this->template Run(); } diff --git a/test/batched_gemm_reduce/CMakeLists.txt b/test/batched_gemm_reduce/CMakeLists.txt index 1a8d7112a3..0710f4647a 100644 --- a/test/batched_gemm_reduce/CMakeLists.txt +++ b/test/batched_gemm_reduce/CMakeLists.txt @@ -1,5 +1,10 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp) target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility) target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance) -endif() + set(target 1) + endif() +endforeach() diff --git a/test/batched_gemm_softmax_gemm/CMakeLists.txt b/test/batched_gemm_softmax_gemm/CMakeLists.txt index dee0bb56ae..984cc3c160 100644 --- a/test/batched_gemm_softmax_gemm/CMakeLists.txt +++ b/test/batched_gemm_softmax_gemm/CMakeLists.txt @@ -1,7 +1,11 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(test_batched_gemm_softmax_gemm) - add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp) target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance) add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt index cb68af06c6..4d04be3faa 100644 --- a/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt +++ b/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(test_batched_gemm_softmax_gemm_permute) add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp) @@ -14,4 +17,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance) add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16) add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/test/contraction/CMakeLists.txt b/test/contraction/CMakeLists.txt index ec44151f53..1f6e0ed341 100644 --- a/test/contraction/CMakeLists.txt +++ b/test/contraction/CMakeLists.txt @@ -1,6 +1,11 @@ add_gtest_executable(test_contraction test_contraction.cpp) target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance) -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_gtest_executable(test_contraction_interface test_contraction_interface.cpp) target_link_libraries(test_contraction_interface PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance) -endif() + set(target 1) + endif() +endforeach() diff --git a/test/convnd_bwd_data/CMakeLists.txt b/test/convnd_bwd_data/CMakeLists.txt index 4bfd21945d..f734b46f53 100644 --- a/test/convnd_bwd_data/CMakeLists.txt +++ b/test/convnd_bwd_data/CMakeLists.txt @@ -1,4 +1,9 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_gtest_executable(test_convnd_bwd_data convnd_bwd_data.cpp) target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance) -endif() \ No newline at end of file + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/test/convnd_fwd/CMakeLists.txt b/test/convnd_fwd/CMakeLists.txt index 058f990266..745aceffc9 100644 --- a/test/convnd_fwd/CMakeLists.txt +++ b/test/convnd_fwd/CMakeLists.txt @@ -1,4 +1,9 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_gtest_executable(test_convnd_fwd convnd_fwd.cpp) target_link_libraries(test_convnd_fwd PRIVATE utility device_conv2d_fwd_instance) -endif() + set(target 1) + endif() +endforeach() diff --git a/test/gemm_layernorm/CMakeLists.txt b/test/gemm_layernorm/CMakeLists.txt index b2a5178ac4..56b8d7737a 100644 --- a/test/gemm_layernorm/CMakeLists.txt +++ b/test/gemm_layernorm/CMakeLists.txt @@ -1,6 +1,11 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(test_gemm_layernorm) add_gtest_executable(test_gemm_add_relu_add_layernorm_fp16 test_gemm_add_relu_add_layernorm_fp16.cpp) target_link_libraries(test_gemm_add_relu_add_layernorm_fp16 PRIVATE utility device_gemm_add_relu_add_layernorm_instance) add_dependencies(test_gemm_layernorm test_gemm_add_relu_add_layernorm_fp16) -endif() + set(target 1) + endif() +endforeach() diff --git a/test/gemm_split_k/CMakeLists.txt b/test/gemm_split_k/CMakeLists.txt index 2274854f88..caf30fca59 100644 --- a/test/gemm_split_k/CMakeLists.txt +++ b/test/gemm_split_k/CMakeLists.txt @@ -1,4 +1,9 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_gtest_executable(test_gemm_splitk test_gemm_splitk.cpp) target_link_libraries(test_gemm_splitk PRIVATE utility device_gemm_splitk_instance) -endif() + set(target 1) + endif() +endforeach() diff --git a/test/grouped_convnd_bwd_weight/CMakeLists.txt b/test/grouped_convnd_bwd_weight/CMakeLists.txt index da554f677a..8be872bc69 100644 --- a/test/grouped_convnd_bwd_weight/CMakeLists.txt +++ b/test/grouped_convnd_bwd_weight/CMakeLists.txt @@ -1,4 +1,9 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_gtest_executable(test_grouped_convnd_bwd_weight grouped_convnd_bwd_weight.cpp) - target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance) -endif() \ No newline at end of file + target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance) + set(target 1) + endif() +endforeach() \ No newline at end of file diff --git a/test/grouped_gemm/CMakeLists.txt b/test/grouped_gemm/CMakeLists.txt index 40f634d8b3..8c57b667e2 100644 --- a/test/grouped_gemm/CMakeLists.txt +++ b/test/grouped_gemm/CMakeLists.txt @@ -1,4 +1,7 @@ -if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") +list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) +set(target 0) +foreach(gpu IN LISTS GPU_TARGETS) + if(gpu IN_LIST gpu_list AND target EQUAL 0) add_custom_target(test_grouped_gemm) add_gtest_executable(test_grouped_gemm_splitk test_grouped_gemm_splitk.cpp) add_gtest_executable(test_grouped_gemm_interface test_grouped_gemm_interface.cpp) @@ -6,4 +9,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M target_link_libraries(test_grouped_gemm_interface PRIVATE utility device_grouped_gemm_instance) add_dependencies(test_grouped_gemm test_grouped_gemm_splitk test_grouped_gemm_interface) -endif() + set(target 1) + endif() +endforeach()