Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "42903973b7083e92e606c95e4bcb4df04ab8a6d1"
Unverified Commit 5d37d7bf authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Reorganize files, Part 1 (#119) 

* delete obselete files

* move files

* build

* update cmake

* update cmake

* fix build

* reorg examples

* update cmake for example and test
parent 245f7414
Show whitespace changes
Inline Side-by-side
Showing with 25 additions and 1754 deletions
CMakeLists.txt
View file @ 5d37d7bf
......@@ -45,7 +45,6 @@ message("OpenMP_gomp_LIBRARY: ${OpenMP_gomp_LIBRARY}")
message("OpenMP_pthread_LIBRARY: ${OpenMP_pthread_LIBRARY}")
message("OpenMP_CXX_FLAGS: ${OpenMP_CXX_FLAGS}")
# set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
link_libraries(${OpenMP_gomp_LIBRARY})
link_libraries(${OpenMP_pthread_LIBRARY})
......@@ -71,17 +70,17 @@ if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
endif()
message(STATUS "Build with HIP ${HIP_VERSION}")
## half
#find_path(HALF_INCLUDE_DIR half.hpp)
set(HALF_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/external/half/include")
message("HALF_INCLUDE_DIR: ${HALF_INCLUDE_DIR}")
rocm_create_package(
NAME CK-${CK_BACKEND}
DESCRIPTION "High Performance Composable Kernels for AMD GPUs"
DESCRIPTION "High Performance Composable Kernel for AMD GPUs"
LDCONFIG
)
## half
set(HALF_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/external/include/half")
message("HALF_INCLUDE_DIR: ${HALF_INCLUDE_DIR}")
## tidy
include(EnableCompilerWarnings)
set(CK_TIDY_ERRORS ERRORS * -readability-inconsistent-declaration-parameter-name)
......@@ -184,7 +183,6 @@ enable_clang_tidy(
-cppcoreguidelines-narrowing-conversions
-altera-struct-pack-align
-cppcoreguidelines-prefer-member-initializer
${CK_TIDY_CHECKS}
${CK_TIDY_ERRORS}
HEADER_FILTER
......@@ -214,69 +212,36 @@ enable_cppcheck(
unmatchedSuppression
FORCE
SOURCES
host/host_tensor/src
host/driver_offline/src
composable_kernel/src/kernel_wrapper
library/src
INCLUDE
host/host_tensor/include
host/device/include
host/solver/include
host/driver_offline/include
composable_kernel/include/*
${CMAKE_CURRENT_SOURCE_DIR}/include
${CMAKE_CURRENT_BINARY_DIR}/include
${CMAKE_CURRENT_SOURCE_DIR}/library/include
DEFINE
CPPCHECK=1
__linux__=1
)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
file(GLOB_RECURSE COMPOSABLE_KERNEL_HEADERS "composable_kernel/include/*/*.hpp")
file(GLOB_RECURSE DEVICE_OPS_HEADERS "device_operation/include/*.hpp")
file(GLOB_RECURSE DEVICE_OPS_SOURCE "device_operation/*.cpp")
set(CK_HEADERS ${COMPOSABLE_KERNEL_HEADERS} ${DEVICE_OPS_HEADERS})
set(CK_SOURCE ${DEVICE_OPS_SOURCE})
add_library(composable_kernel ${CK_SOURCE})
configure_file("${PROJECT_SOURCE_DIR}/include/ck/hip_version.hpp.in" "${PROJECT_BINARY_DIR}/include/ck/hip_version.hpp")
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/composable_kernel/include>
)
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/device_operation/include>
)
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/host/include>
)
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/host/host_tensor/include>
)
# The following should eventually be removed
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/composable_kernel/include/utility>
)
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_operation>
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include
${PROJECT_BINARY_DIR}/include
${PROJECT_SOURCE_DIR}/library/include
)
target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description>
)
# clang_tidy_check(composable_kernel)
SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV")
if(BUILD_DEV)
target_compile_options(composable_kernel PRIVATE -Werror)
target_compile_options(composable_kernel PRIVATE -Weverything)
add_compile_options(-Werror)
add_compile_options(-Weverything)
endif()
message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/hip_version.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/hip_version.hpp")
add_subdirectory(host)
add_subdirectory(device_operation)
add_subdirectory(library)
add_subdirectory(example)
add_subdirectory(profiler)
add_subdirectory(test)
add_subdirectory(profiler)
composable_kernel/include/gridwise_operation_wrapper.hpp deleted 100644 → 0
View file @ 245f7414
#ifndef CK_GRIDWISE_OPERATION_KERNEL_WRAPPER
#define CK_GRIDWISE_OPERATION_KERNEL_WRAPPER
template <typename GridwiseOp, typename... Xs>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
run_gridwise_operation(Xs... xs)
{
GridwiseOp{}.Run(xs...);
}
#endif
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v4r4_dlops_nchw_kcyx_nkhw.cpp deleted 100644 → 0
View file @ 245f7414
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_dlops_v1r2.hpp"
#include "transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp"
using namespace ck;
constexpr DataTypeEnum_t ABDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_ABDataTypeEnum);
constexpr DataTypeEnum_t AccDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_AccDataTypeEnum);
constexpr DataTypeEnum_t CDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_CDataTypeEnum);
using FloatAB = typename get_datatype_from_enum<ABDataTypeEnum>::type;
using FloatAcc = typename get_datatype_from_enum<AccDataTypeEnum>::type;
using FloatC = typename get_datatype_from_enum<CDataTypeEnum>::type;
constexpr index_t BlockSize = CK_PARAM_BlockSize;
constexpr index_t MPerBlock = CK_PARAM_MPerBlock;
constexpr index_t NPerBlock = CK_PARAM_NPerBlock;
constexpr index_t KPerBlock = CK_PARAM_KPerBlock;
constexpr index_t M1PerThread = CK_PARAM_M1PerThread;
constexpr index_t N1PerThread = CK_PARAM_N1PerThread;
constexpr index_t KPerThread = CK_PARAM_KPerThread;
constexpr index_t M1N1ThreadClusterM10 = CK_PARAM_M1N1ThreadClusterM10;
constexpr index_t M1N1ThreadClusterN10 = CK_PARAM_M1N1ThreadClusterN10;
constexpr index_t M1N1ThreadClusterM11 = CK_PARAM_M1N1ThreadClusterM11;
constexpr index_t M1N1ThreadClusterN11 = CK_PARAM_M1N1ThreadClusterN11;
using ABlockTransferThreadSliceLengths_K_M0_M1 =
Sequence<CK_PARAM_ABlockTransferThreadSliceLengths_K_M0_M1>;
using ABlockTransferThreadClusterLengths_K_M0_M1 =
Sequence<CK_PARAM_ABlockTransferThreadClusterLengths_K_M0_M1>;
using ABlockTransferThreadClusterArrangeOrder =
Sequence<CK_PARAM_ABlockTransferThreadClusterArrangeOrder>;
using ABlockTransferSrcAccessOrder = Sequence<CK_PARAM_ABlockTransferSrcAccessOrder>;
constexpr index_t ABlockTransferSrcVectorDim = CK_PARAM_ABlockTransferSrcVectorDim;
constexpr index_t ABlockTransferSrcScalarPerVector = CK_PARAM_ABlockTransferSrcScalarPerVector;
constexpr index_t ABlockTransferDstScalarPerVector_M1 =
CK_PARAM_ABlockTransferDstScalarPerVector_M1;
constexpr bool AThreadTransferSrcResetCoordinateAfterRun =
static_cast<bool>(CK_PARAM_AThreadTransferSrcResetCoordinateAfterRun);
using BBlockTransferThreadSliceLengths_K_N0_N1 =
Sequence<CK_PARAM_BBlockTransferThreadSliceLengths_K_N0_N1>;
using BBlockTransferThreadClusterLengths_K_N0_N1 =
Sequence<CK_PARAM_BBlockTransferThreadClusterLengths_K_N0_N1>;
using BBlockTransferThreadClusterArrangeOrder =
Sequence<CK_PARAM_BBlockTransferThreadClusterArrangeOrder>;
using BBlockTransferSrcAccessOrder = Sequence<CK_PARAM_BBlockTransferSrcAccessOrder>;
constexpr index_t BBlockTransferSrcVectorDim = CK_PARAM_BBlockTransferSrcVectorDim;
constexpr index_t BBlockTransferSrcScalarPerVector = CK_PARAM_BBlockTransferSrcScalarPerVector;
constexpr index_t BBlockTransferDstScalarPerVector_N1 =
CK_PARAM_BBlockTransferDstScalarPerVector_N1;
constexpr bool BThreadTransferSrcResetCoordinateAfterRun =
static_cast<bool>(CK_PARAM_BThreadTransferSrcResetCoordinateAfterRun);
using CThreadTransferSrcDstAccessOrder = Sequence<CK_PARAM_CThreadTransferSrcDstAccessOrder>;
constexpr index_t CThreadTransferSrcDstVectorDim = CK_PARAM_CThreadTransferSrcDstVectorDim;
constexpr index_t CThreadTransferDstScalarPerVector = CK_PARAM_CThreadTransferDstScalarPerVector;
constexpr bool HasMainKBlockLoop = static_cast<bool>(CK_PARAM_HAS_MAIN_KBLOCK_LOOP);
constexpr bool HasDoubleTailKBlockLoop = static_cast<bool>(CK_PARAM_HAS_DOUBLE_TAIL_KBLOCK_LOOP);
extern "C" __global__ void convolution_forward_implicit_gemm_v4r4_dlops_nchw_kcyx_nkhw_prepare(
int n,
int c,
int hi,
int wi,
int k,
int y,
int x,
int convStrideH,
int convStrideW,
int convDilationY,
int convDilationX,
int leftPadH,
int leftPadW,
int rightPadH,
int rightPadW,
void* p_a_k_m0_m1_grid_desc,
void* p_b_k_n0_n1_grid_desc,
void* p_c_m0_m10_m11_n0_n10_n11_grid_desc,
void* p_cblockid_to_m0_n0_block_cluster_adaptor)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
const index_t ho = (hi + leftPadH + rightPadH - convDilationY * (y - 1) - 1) / convStrideH + 1;
const index_t wo = (wi + leftPadW + rightPadW - convDilationX * (x - 1) - 1) / convStrideW + 1;
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(make_tuple(n, c, hi, wi));
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(make_tuple(k, c, y, x));
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(make_tuple(n, k, ho, wo));
const auto descs = transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_pad(
wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
make_tuple(convStrideH, convStrideW),
make_tuple(convDilationY, convDilationX),
make_tuple(leftPadH, leftPadW),
make_tuple(rightPadH, rightPadW));
const auto a_k_m_grid_desc = descs[I0];
const auto b_k_n_grid_desc = descs[I1];
const auto c_m_n_grid_desc = descs[I2];
using AKMGridDesc = decltype(a_k_m_grid_desc);
using BKNGridDesc = decltype(b_k_n_grid_desc);
using CMNGridDesc = decltype(c_m_n_grid_desc);
using AGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{})));
using BGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})));
using CGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})));
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0>;
using BGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>;
using GridwiseGemm =
GridwiseGemmDlops_km_kn_mn_v1r2<BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set, /* ToDo tunable */
AKMGridDesc,
BKNGridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
M1PerThread,
N1PerThread,
KPerThread,
M1N1ThreadClusterM10,
M1N1ThreadClusterN10,
M1N1ThreadClusterM11,
M1N1ThreadClusterN11,
ABlockTransferThreadSliceLengths_K_M0_M1,
ABlockTransferThreadClusterLengths_K_M0_M1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_M1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K_N0_N1,
BBlockTransferThreadClusterLengths_K_N0_N1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_N1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks>;
auto a_k_m0_m1_grid_desc = GridwiseGemm::MakeAKM0M1GridDescriptor(a_k_m_grid_desc);
auto b_k_n0_n1_grid_desc = GridwiseGemm::MakeBKN0N1GridDescriptor(b_k_n_grid_desc);
auto c_m0_m10_m11_n0_n10_n11_grid_desc =
GridwiseGemm::MakeCM0M10M11N0N10N11GridDescriptor(c_m_n_grid_desc);
auto cblockid_to_m0_n0_block_cluster_adaptor =
GridwiseGemm::MakeCBlockIdToM0N0BlockClusterAdaptor(c_m_n_grid_desc);
if(hipThreadIdx_x == 0)
{
*static_cast<decltype(a_k_m0_m1_grid_desc)*>(p_a_k_m0_m1_grid_desc) = a_k_m0_m1_grid_desc;
*static_cast<decltype(b_k_n0_n1_grid_desc)*>(p_b_k_n0_n1_grid_desc) = b_k_n0_n1_grid_desc;
*static_cast<decltype(c_m0_m10_m11_n0_n10_n11_grid_desc)*>(
p_c_m0_m10_m11_n0_n10_n11_grid_desc) = c_m0_m10_m11_n0_n10_n11_grid_desc;
*static_cast<decltype(cblockid_to_m0_n0_block_cluster_adaptor)*>(
p_cblockid_to_m0_n0_block_cluster_adaptor) = cblockid_to_m0_n0_block_cluster_adaptor;
};
};
extern "C" __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
convolution_forward_implicit_gemm_v4r4_dlops_nchw_kcyx_nkhw(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_k_m0_m1_grid_desc,
const void CONSTANT* p_b_k_n0_n1_grid_desc,
const void CONSTANT* p_c_m0_m10_m11_n0_n10_n11_grid_desc,
const void CONSTANT* p_cblockid_to_m0_n0_block_cluster_adaptor)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto in_n_c_hi_wi_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 28, 28));
constexpr auto wei_k_c_y_x_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 3, 3));
constexpr auto out_n_k_ho_wo_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 28, 28));
constexpr auto descs =
transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_pad(wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1));
constexpr auto a_k_m_grid_desc = descs[I0];
constexpr auto b_k_n_grid_desc = descs[I1];
constexpr auto c_m_n_grid_desc = descs[I2];
using AKMGridDesc = decltype(a_k_m_grid_desc);
using BKNGridDesc = decltype(b_k_n_grid_desc);
using CMNGridDesc = decltype(c_m_n_grid_desc);
using AGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{})));
using BGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})));
using CGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})));
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0>;
using BGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>;
using GridwiseGemm =
GridwiseGemmDlops_km_kn_mn_v1r2<BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set, /* ToDo tunable */
AKMGridDesc,
BKNGridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
M1PerThread,
N1PerThread,
KPerThread,
M1N1ThreadClusterM10,
M1N1ThreadClusterN10,
M1N1ThreadClusterM11,
M1N1ThreadClusterN11,
ABlockTransferThreadSliceLengths_K_M0_M1,
ABlockTransferThreadClusterLengths_K_M0_M1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_M1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K_N0_N1,
BBlockTransferThreadClusterLengths_K_N0_N1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_N1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks>;
constexpr auto a_k_m0_m1_grid_desc_tmp =
GridwiseGemm::MakeAKM0M1GridDescriptor(a_k_m_grid_desc);
constexpr auto b_k_n0_n1_grid_desc_tmp =
GridwiseGemm::MakeBKN0N1GridDescriptor(b_k_n_grid_desc);
constexpr auto c_m0_m10_m11_n0_n10_n11_grid_desc_tmp =
GridwiseGemm::MakeCM0M10M11N0N10N11GridDescriptor(c_m_n_grid_desc);
constexpr auto cblockid_to_m0_n0_block_cluster_adaptor_tmp =
GridwiseGemm::MakeCBlockIdToM0N0BlockClusterAdaptor(c_m_n_grid_desc);
using AKM0M1GridDesc = decltype(a_k_m0_m1_grid_desc_tmp);
using BKN0N1GridDesc = decltype(b_k_n0_n1_grid_desc_tmp);
using CM0M10M11N0N10N11GridDesc = decltype(c_m0_m10_m11_n0_n10_n11_grid_desc_tmp);
using CBlockIdToM0N0BlockClusterAdaptor = decltype(cblockid_to_m0_n0_block_cluster_adaptor_tmp);
const auto a_k_m0_m1_grid_desc =
*reinterpret_cast<const AKM0M1GridDesc*>((const void*)p_a_k_m0_m1_grid_desc);
const auto b_k_n0_n1_grid_desc =
*reinterpret_cast<const BKN0N1GridDesc*>((const void*)p_b_k_n0_n1_grid_desc);
const auto c_m0_m10_m11_n0_n10_n11_grid_desc =
*reinterpret_cast<const CM0M10M11N0N10N11GridDesc*>(
(const void*)p_c_m0_m10_m11_n0_n10_n11_grid_desc);
const auto cblockid_to_m0_n0_block_cluster_adaptor =
*reinterpret_cast<const CBlockIdToM0N0BlockClusterAdaptor*>(
(const void*)p_cblockid_to_m0_n0_block_cluster_adaptor);
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::Run(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_block,
a_k_m0_m1_grid_desc,
b_k_n0_n1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor,
integral_constant<bool, HasMainKBlockLoop>{},
integral_constant<bool, HasDoubleTailKBlockLoop>{});
};
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw.cpp deleted 100644 → 0
View file @ 245f7414
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r3.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
using namespace ck;
constexpr DataTypeEnum_t ABDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_ABDataTypeEnum);
constexpr DataTypeEnum_t AccDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_AccDataTypeEnum);
constexpr DataTypeEnum_t CDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_CDataTypeEnum);
using FloatAB = typename get_datatype_from_enum<ABDataTypeEnum>::type;
using FloatAcc = typename get_datatype_from_enum<AccDataTypeEnum>::type;
using FloatC = typename get_datatype_from_enum<CDataTypeEnum>::type;
constexpr index_t BlockSize = CK_PARAM_BlockSize;
constexpr index_t MPerBlock = CK_PARAM_MPerBlock;
constexpr index_t NPerBlock = CK_PARAM_NPerBlock;
constexpr index_t KPerBlock = CK_PARAM_KPerBlock;
constexpr index_t MPerWave = CK_PARAM_MPerWave;
constexpr index_t NPerWave = CK_PARAM_NPerWave;
constexpr index_t MRepeat = CK_PARAM_MRepeat;
constexpr index_t NRepeat = CK_PARAM_NRepeat;
constexpr index_t K1 = CK_PARAM_K1;
using ABlockTransferThreadSliceLengths_K0_M_K1 =
Sequence<CK_PARAM_ABlockTransferThreadSliceLengths_K0_M_K1>;
using ABlockTransferThreadClusterLengths_K0_M_K1 =
Sequence<CK_PARAM_ABlockTransferThreadClusterLengths_K0_M_K1>;
using ABlockTransferThreadClusterArrangeOrder =
Sequence<CK_PARAM_ABlockTransferThreadClusterArrangeOrder>;
using ABlockTransferSrcAccessOrder = Sequence<CK_PARAM_ABlockTransferSrcAccessOrder>;
constexpr index_t ABlockTransferSrcVectorDim = CK_PARAM_ABlockTransferSrcVectorDim;
constexpr index_t ABlockTransferSrcScalarPerVector = CK_PARAM_ABlockTransferSrcScalarPerVector;
constexpr index_t ABlockTransferDstScalarPerVector_K1 =
CK_PARAM_ABlockTransferDstScalarPerVector_K1;
constexpr bool AThreadTransferSrcResetCoordinateAfterRun =
static_cast<bool>(CK_PARAM_AThreadTransferSrcResetCoordinateAfterRun);
using BBlockTransferThreadSliceLengths_K0_N_K1 =
Sequence<CK_PARAM_BBlockTransferThreadSliceLengths_K0_N_K1>;
using BBlockTransferThreadClusterLengths_K0_N_K1 =
Sequence<CK_PARAM_BBlockTransferThreadClusterLengths_K0_N_K1>;
using BBlockTransferThreadClusterArrangeOrder =
Sequence<CK_PARAM_BBlockTransferThreadClusterArrangeOrder>;
using BBlockTransferSrcAccessOrder = Sequence<CK_PARAM_BBlockTransferSrcAccessOrder>;
constexpr index_t BBlockTransferSrcVectorDim = CK_PARAM_BBlockTransferSrcVectorDim;
constexpr index_t BBlockTransferSrcScalarPerVector = CK_PARAM_BBlockTransferSrcScalarPerVector;
constexpr index_t BBlockTransferDstScalarPerVector_K1 =
CK_PARAM_BBlockTransferDstScalarPerVector_K1;
constexpr bool BThreadTransferSrcResetCoordinateAfterRun =
static_cast<bool>(CK_PARAM_BThreadTransferSrcResetCoordinateAfterRun);
using CThreadTransferSrcDstAccessOrder = Sequence<CK_PARAM_CThreadTransferSrcDstAccessOrder>;
constexpr index_t CThreadTransferSrcDstVectorDim = CK_PARAM_CThreadTransferSrcDstVectorDim;
constexpr index_t CThreadTransferDstScalarPerVector = CK_PARAM_CThreadTransferDstScalarPerVector;
extern "C" __global__ void convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw_prepare(
int n,
int c,
int hi,
int wi,
int k,
int y,
int x,
int convStrideH,
int convStrideW,
int convDilationY,
int convDilationX,
int leftPadH,
int leftPadW,
int rightPadH,
int rightPadW,
void* p_a_k0_m_k1_grid_desc,
void* p_b_k0_n_k1_grid_desc,
void* p_c_m0_m1_m2_n_grid_desc,
void* p_cblockid_to_m0_n0_block_cluster_adaptor)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
const index_t ho = (hi + leftPadH + rightPadH - convDilationY * (y - 1) - 1) / convStrideH + 1;
const index_t wo = (wi + leftPadW + rightPadW - convDilationX * (x - 1) - 1) / convStrideW + 1;
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(make_tuple(n, c, hi, wi));
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(make_tuple(k, c, y, x));
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(make_tuple(n, k, ho, wo));
const auto descs = transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(
wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
make_tuple(convStrideH, convStrideW),
make_tuple(convDilationY, convDilationX),
make_tuple(leftPadH, leftPadW),
make_tuple(rightPadH, rightPadW),
Number<K1>{});
const auto a_k0_m_k1_grid_desc = descs[I0];
const auto b_k0_n_k1_grid_desc = descs[I1];
const auto c_m_n_grid_desc = descs[I2];
using AK0MK1GridDesc = decltype(a_k0_m_k1_grid_desc);
using BK0NK1GridDesc = decltype(b_k0_n_k1_grid_desc);
using CMNGridDesc = decltype(c_m_n_grid_desc);
using AGridStepHacks = decltype(make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{})));
using BGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})));
using CGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})));
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0>;
using BGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>;
using GridwiseGemm =
GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
AK0MK1GridDesc,
BK0NK1GridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
MPerWave,
NPerWave,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K0_N_K1,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
false>;
auto c_m0_m1_m2_n_grid_desc = GridwiseGemm::MakeCM0M1M2NGridDescriptor(c_m_n_grid_desc);
auto cblockid_to_m0_n0_block_cluster_adaptor =
GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc);
if(hipThreadIdx_x == 0)
{
*static_cast<remove_cv_t<decltype(a_k0_m_k1_grid_desc)>*>(p_a_k0_m_k1_grid_desc) =
a_k0_m_k1_grid_desc;
*static_cast<remove_cv_t<decltype(b_k0_n_k1_grid_desc)>*>(p_b_k0_n_k1_grid_desc) =
b_k0_n_k1_grid_desc;
*static_cast<decltype(c_m0_m1_m2_n_grid_desc)*>(p_c_m0_m1_m2_n_grid_desc) =
c_m0_m1_m2_n_grid_desc;
*static_cast<decltype(cblockid_to_m0_n0_block_cluster_adaptor)*>(
p_cblockid_to_m0_n0_block_cluster_adaptor) = cblockid_to_m0_n0_block_cluster_adaptor;
}
};
extern "C" __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_k0_m_k1_grid_desc,
const void CONSTANT* p_b_k0_n_k1_grid_desc,
const void CONSTANT* p_c_m0_m1_m2_n_grid_desc,
const void CONSTANT* p_cblockid_to_m0_n0_block_cluster_adaptor)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto in_n_c_hi_wi_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 28, 28));
constexpr auto wei_k_c_y_x_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 3, 3));
constexpr auto out_n_k_ho_wo_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 28, 28));
constexpr auto descs =
transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
Number<K1>{});
constexpr auto a_k0_m_k1_grid_desc_tmp = descs[I0];
constexpr auto b_k0_n_k1_grid_desc_tmp = descs[I1];
constexpr auto c_m_n_grid_desc = descs[I2];
using AGridStepHacks = decltype(make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{})));
using BGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})));
using CGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})));
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0>;
using BGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>;
using AK0MK1GridDesc = decltype(a_k0_m_k1_grid_desc_tmp);
using BK0NK1GridDesc = decltype(b_k0_n_k1_grid_desc_tmp);
using CMNGridDesc = decltype(c_m_n_grid_desc);
using GridwiseGemm =
GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
AK0MK1GridDesc,
BK0NK1GridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
MPerWave,
NPerWave,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K0_N_K1,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
false>;
constexpr auto c_m0_m1_m2_n_grid_desc_tmp =
GridwiseGemm::MakeCM0M1M2NGridDescriptor(c_m_n_grid_desc);
constexpr auto cblockid_to_m0_n0_block_cluster_adaptor_tmp =
GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc);
using CM0M1M2NGridDesc = decltype(c_m0_m1_m2_n_grid_desc_tmp);
using CBlockIdToM0N0BlockClusterAdaptor = decltype(cblockid_to_m0_n0_block_cluster_adaptor_tmp);
const auto a_k0_m_k1_grid_desc =
*reinterpret_cast<const AK0MK1GridDesc*>((const void*)p_a_k0_m_k1_grid_desc);
const auto b_k0_n_k1_grid_desc =
*reinterpret_cast<const BK0NK1GridDesc*>((const void*)p_b_k0_n_k1_grid_desc);
const auto c_m0_m1_m2_n_grid_desc =
*reinterpret_cast<const CM0M1M2NGridDesc*>((const void*)p_c_m0_m1_m2_n_grid_desc);
const auto cblockid_to_m0_n0_block_cluster_adaptor =
*reinterpret_cast<const CBlockIdToM0N0BlockClusterAdaptor*>(
(const void*)p_cblockid_to_m0_n0_block_cluster_adaptor);
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::Run(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_block,
a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc,
c_m0_m1_m2_n_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
};
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v4r4_xdlops_nhwc_kyxc_nhwk.cpp deleted 100644 → 0
View file @ 245f7414
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r3.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp"
using namespace ck;
constexpr DataTypeEnum_t ABDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_ABDataTypeEnum);
constexpr DataTypeEnum_t AccDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_AccDataTypeEnum);
constexpr DataTypeEnum_t CDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_CDataTypeEnum);
using FloatAB = typename get_datatype_from_enum<ABDataTypeEnum>::type;
using FloatAcc = typename get_datatype_from_enum<AccDataTypeEnum>::type;
using FloatC = typename get_datatype_from_enum<CDataTypeEnum>::type;
constexpr index_t BlockSize = CK_PARAM_BlockSize;
constexpr index_t MPerBlock = CK_PARAM_MPerBlock;
constexpr index_t NPerBlock = CK_PARAM_NPerBlock;
constexpr index_t KPerBlock = CK_PARAM_KPerBlock;
constexpr index_t MPerWave = CK_PARAM_MPerWave;
constexpr index_t NPerWave = CK_PARAM_NPerWave;
constexpr index_t MRepeat = CK_PARAM_MRepeat;
constexpr index_t NRepeat = CK_PARAM_NRepeat;
constexpr index_t K1 = CK_PARAM_K1;
using ABlockTransferThreadSliceLengths_K0_M_K1 =
Sequence<CK_PARAM_ABlockTransferThreadSliceLengths_K0_M_K1>;
using ABlockTransferThreadClusterLengths_K0_M_K1 =
Sequence<CK_PARAM_ABlockTransferThreadClusterLengths_K0_M_K1>;
using ABlockTransferThreadClusterArrangeOrder =
Sequence<CK_PARAM_ABlockTransferThreadClusterArrangeOrder>;
using ABlockTransferSrcAccessOrder = Sequence<CK_PARAM_ABlockTransferSrcAccessOrder>;
constexpr index_t ABlockTransferSrcVectorDim = CK_PARAM_ABlockTransferSrcVectorDim;
constexpr index_t ABlockTransferSrcScalarPerVector = CK_PARAM_ABlockTransferSrcScalarPerVector;
constexpr index_t ABlockTransferDstScalarPerVector_K1 =
CK_PARAM_ABlockTransferDstScalarPerVector_K1;
constexpr bool AThreadTransferSrcResetCoordinateAfterRun =
static_cast<bool>(CK_PARAM_AThreadTransferSrcResetCoordinateAfterRun);
using BBlockTransferThreadSliceLengths_K0_N_K1 =
Sequence<CK_PARAM_BBlockTransferThreadSliceLengths_K0_N_K1>;
using BBlockTransferThreadClusterLengths_K0_N_K1 =
Sequence<CK_PARAM_BBlockTransferThreadClusterLengths_K0_N_K1>;
using BBlockTransferThreadClusterArrangeOrder =
Sequence<CK_PARAM_BBlockTransferThreadClusterArrangeOrder>;
using BBlockTransferSrcAccessOrder = Sequence<CK_PARAM_BBlockTransferSrcAccessOrder>;
constexpr index_t BBlockTransferSrcVectorDim = CK_PARAM_BBlockTransferSrcVectorDim;
constexpr index_t BBlockTransferSrcScalarPerVector = CK_PARAM_BBlockTransferSrcScalarPerVector;
constexpr index_t BBlockTransferDstScalarPerVector_K1 =
CK_PARAM_BBlockTransferDstScalarPerVector_K1;
constexpr bool BThreadTransferSrcResetCoordinateAfterRun =
static_cast<bool>(CK_PARAM_BThreadTransferSrcResetCoordinateAfterRun);
using CThreadTransferSrcDstAccessOrder = Sequence<CK_PARAM_CThreadTransferSrcDstAccessOrder>;
constexpr index_t CThreadTransferSrcDstVectorDim = CK_PARAM_CThreadTransferSrcDstVectorDim;
constexpr index_t CThreadTransferDstScalarPerVector = CK_PARAM_CThreadTransferDstScalarPerVector;
extern "C" __global__ void convolution_forward_implicit_gemm_v4r4_xdlops_nhwc_kyxc_nhwk_prepare(
int n,
int hi,
int wi,
int c,
int k,
int y,
int x,
int convStrideH,
int convStrideW,
int convDilationY,
int convDilationX,
int leftPadH,
int leftPadW,
int rightPadH,
int rightPadW,
void* p_a_k0_m_k1_grid_desc,
void* p_b_k0_n_k1_grid_desc,
void* p_c_m0_m1_m2_n_grid_desc,
void* p_cblockid_to_m0_n0_block_cluster_adaptor)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
const index_t ho = (hi + leftPadH + rightPadH - convDilationY * (y - 1) - 1) / convStrideH + 1;
const index_t wo = (wi + leftPadW + rightPadW - convDilationX * (x - 1) - 1) / convStrideW + 1;
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(make_tuple(n, hi, wi, c));
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(make_tuple(k, y, x, c));
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(make_tuple(n, ho, wo, k));
const auto descs = transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(
in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
make_tuple(convStrideH, convStrideW),
make_tuple(convDilationY, convDilationX),
make_tuple(leftPadH, leftPadW),
make_tuple(rightPadH, rightPadW),
Number<K1>{});
const auto a_k0_m_k1_grid_desc = descs[I0];
const auto b_k0_n_k1_grid_desc = descs[I1];
const auto c_m_n_grid_desc = descs[I2];
using AK0MK1GridDesc = decltype(a_k0_m_k1_grid_desc);
using BK0NK1GridDesc = decltype(b_k0_n_k1_grid_desc);
using CMNGridDesc = decltype(c_m_n_grid_desc);
using BGridStepHacks = decltype(make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{})));
using AGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})));
using CGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})));
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>;
using BGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0>;
using GridwiseGemm =
GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
AK0MK1GridDesc,
BK0NK1GridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
MPerWave,
NPerWave,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K0_N_K1,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
false>;
auto c_m0_m1_m2_n_grid_desc = GridwiseGemm::MakeCM0M1M2NGridDescriptor(c_m_n_grid_desc);
auto cblockid_to_m0_n0_block_cluster_adaptor =
GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc);
if(hipThreadIdx_x == 0)
{
*static_cast<remove_cv_t<decltype(a_k0_m_k1_grid_desc)>*>(p_a_k0_m_k1_grid_desc) =
a_k0_m_k1_grid_desc;
*static_cast<remove_cv_t<decltype(b_k0_n_k1_grid_desc)>*>(p_b_k0_n_k1_grid_desc) =
b_k0_n_k1_grid_desc;
*static_cast<decltype(c_m0_m1_m2_n_grid_desc)*>(p_c_m0_m1_m2_n_grid_desc) =
c_m0_m1_m2_n_grid_desc;
*static_cast<decltype(cblockid_to_m0_n0_block_cluster_adaptor)*>(
p_cblockid_to_m0_n0_block_cluster_adaptor) = cblockid_to_m0_n0_block_cluster_adaptor;
}
};
extern "C" __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
convolution_forward_implicit_gemm_v4r4_xdlops_nhwc_kyxc_nhwk(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_k0_m_k1_grid_desc,
const void CONSTANT* p_b_k0_n_k1_grid_desc,
const void CONSTANT* p_c_m0_m1_m2_n_grid_desc,
const void CONSTANT* p_cblockid_to_m0_n0_block_cluster_adaptor)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto in_n_hi_wi_c_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 28, 28, 256));
constexpr auto wei_k_y_x_c_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 3, 3, 256));
constexpr auto out_n_ho_wo_k_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 28, 28, 256));
constexpr auto descs =
transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
Number<K1>{});
constexpr auto a_k0_m_k1_grid_desc_tmp = descs[I0];
constexpr auto b_k0_n_k1_grid_desc_tmp = descs[I1];
constexpr auto c_m_n_grid_desc = descs[I2];
using AK0MK1GridDesc = decltype(a_k0_m_k1_grid_desc_tmp);
using BK0NK1GridDesc = decltype(b_k0_n_k1_grid_desc_tmp);
using CMNGridDesc = decltype(c_m_n_grid_desc);
using BGridStepHacks = decltype(make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{})));
using AGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})));
using CGridStepHacks = decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})));
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>;
using BGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0>;
using GridwiseGemm =
GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
AK0MK1GridDesc,
BK0NK1GridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
MPerWave,
NPerWave,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K0_N_K1,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
false>;
constexpr auto c_m0_m1_m2_n_grid_desc_tmp =
GridwiseGemm::MakeCM0M1M2NGridDescriptor(c_m_n_grid_desc);
constexpr auto cblockid_to_m0_n0_block_cluster_adaptor_tmp =
GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc);
using CM0M1M2NGridDesc = decltype(c_m0_m1_m2_n_grid_desc_tmp);
using CBlockIdToM0N0BlockClusterAdaptor = decltype(cblockid_to_m0_n0_block_cluster_adaptor_tmp);
const auto a_k0_m_k1_grid_desc =
*reinterpret_cast<const AK0MK1GridDesc*>((const void*)p_a_k0_m_k1_grid_desc);
const auto b_k0_n_k1_grid_desc =
*reinterpret_cast<const BK0NK1GridDesc*>((const void*)p_b_k0_n_k1_grid_desc);
const auto c_m0_m1_m2_n_grid_desc =
*reinterpret_cast<const CM0M1M2NGridDesc*>((const void*)p_c_m0_m1_m2_n_grid_desc);
const auto cblockid_to_m0_n0_block_cluster_adaptor =
*reinterpret_cast<const CBlockIdToM0N0BlockClusterAdaptor*>(
(const void*)p_cblockid_to_m0_n0_block_cluster_adaptor);
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::Run(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_block,
a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc,
c_m0_m1_m2_n_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
};
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw.cpp deleted 100644 → 0
View file @ 245f7414
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_contraction_dlops_v1r2.hpp"
#include "transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp"
using namespace ck;
constexpr DataTypeEnum_t ABDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_ABDataTypeEnum);
constexpr DataTypeEnum_t AccDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_AccDataTypeEnum);
constexpr DataTypeEnum_t CDataTypeEnum = static_cast<DataTypeEnum_t>(CK_PARAM_CDataTypeEnum);
using FloatAB = typename get_datatype_from_enum<ABDataTypeEnum>::type;
using FloatAcc = typename get_datatype_from_enum<AccDataTypeEnum>::type;
using FloatC = typename get_datatype_from_enum<CDataTypeEnum>::type;
constexpr index_t BlockSize = CK_PARAM_BlockSize;
constexpr auto GN0 = Number<CK_PARAM_GN0>{};
constexpr auto GK1 = Number<CK_PARAM_GK1>{};
constexpr index_t GM1PerBlockGM11 = CK_PARAM_GM1PerBlockGM11;
constexpr index_t GN1PerBlockGN11 = CK_PARAM_GN1PerBlockGN11;
constexpr index_t GK0PerBlock = CK_PARAM_GK0PerBlock;
constexpr index_t BM1PerThreadBM11 = CK_PARAM_BM1PerThreadBM11;
constexpr index_t BN1PerThreadBN11 = CK_PARAM_BN1PerThreadBN11;
constexpr index_t BK0PerThread = CK_PARAM_BK0PerThread;
using BM10BN10ThreadClusterBM10Xs = Sequence<CK_PARAM_BM10BN10ThreadClusterBM10Xs>;
using BM10BN10ThreadClusterBN10Xs = Sequence<CK_PARAM_BM10BN10ThreadClusterBN10Xs>;
using ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1 =
Sequence<CK_PARAM_ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1>;
using ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1 =
Sequence<CK_PARAM_ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1>;
using ABlockTransferThreadClusterArrangeOrder = Sequence<1, 2, 3, 0, 4>;
using ABlockTransferSrcAccessOrder = Sequence<3, 2, 1, 0, 4>;
using ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1 =
Sequence<CK_PARAM_ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1>;
using ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1 =
Sequence<CK_PARAM_ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1>;
using ABlockTransferSrcVectorTensorContiguousDimOrder = Sequence<0, 1, 2, 3, 4>;
using BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1 =
Sequence<CK_PARAM_BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1>;
using BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1 =
Sequence<CK_PARAM_BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1>;
using BBlockTransferThreadClusterArrangeOrder = Sequence<0, 4, 1, 2, 3>;
using BBlockTransferSrcAccessOrder = Sequence<4, 3, 2, 0, 1>;
using BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1 =
Sequence<CK_PARAM_BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1>;
using BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1 =
Sequence<CK_PARAM_BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1>;
using BBlockTransferSrcVectorTensorContiguousDimOrder = Sequence<0, 1, 2, 3, 4>;
using CThreadTransferSrcDstAccessOrder = Sequence<3, 4, 5, 0, 1, 2>;
constexpr index_t CThreadTransferSrcDstVectorDim = 5;
constexpr index_t CThreadTransferDstScalarPerVector = CK_PARAM_CThreadTransferDstScalarPerVector;
constexpr bool HasMainKBlockLoop = static_cast<bool>(CK_PARAM_HasMainKBlockLoop);
constexpr bool HasDoubleTailKBlockLoop = static_cast<bool>(CK_PARAM_HasDoubleTailKBlockLoop);
extern "C" __global__ void
convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw_prepare(int N_,
int C_,
int Hi_,
int Wi_,
int K_,
int Y_,
int X_,
int ConvStrideH_,
int ConvStrideW_,
int ConvDilationH_,
int ConvDilationW_,
int InLeftPadH_,
int InLeftPadW_,
int InRightPadH_,
int InRightPadW_,
void* p_desc_tuple)
{
index_t N = static_cast<index_t>(N_);
index_t C = static_cast<index_t>(C_);
index_t Hi = static_cast<index_t>(Hi_);
index_t Wi = static_cast<index_t>(Wi_);
index_t K = static_cast<index_t>(K_);
index_t Y = static_cast<index_t>(Y_);
index_t X = static_cast<index_t>(X_);
index_t ConvStrideH = static_cast<index_t>(ConvStrideH_);
index_t ConvStrideW = static_cast<index_t>(ConvStrideW_);
index_t ConvDilationH = static_cast<index_t>(ConvDilationH_);
index_t ConvDilationW = static_cast<index_t>(ConvDilationW_);
index_t InLeftPadH = static_cast<index_t>(InLeftPadH_);
index_t InLeftPadW = static_cast<index_t>(InLeftPadW_);
index_t InRightPadH = static_cast<index_t>(InRightPadH_);
index_t InRightPadW = static_cast<index_t>(InRightPadW_);
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
const index_t Ho =
(Hi + InLeftPadH + InRightPadH - ConvDilationH * (Y - 1) - 1) / ConvStrideH + 1;
const index_t Wo =
(Wi + InLeftPadW + InRightPadW - ConvDilationW * (X - 1) - 1) / ConvStrideW + 1;
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(make_tuple(N, C, Hi, Wi));
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(make_tuple(K, C, Y, X));
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho, Wo));
const auto descs = transform_forward_convolution_into_contraction_v6r1_nchw_kcyx_nkhw_pad(
wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
make_tuple(ConvStrideH, ConvStrideW),
make_tuple(ConvDilationH, ConvDilationW),
make_tuple(InLeftPadH, InLeftPadW),
make_tuple(InRightPadH, InRightPadW),
GN0,
GK1);
const auto a_grid_desc_gk0_gm0_gm1_gk1 = descs[I0];
const auto b_grid_desc_gk0_gn0_gn1_gk1 = descs[I1];
const auto c_grid_desc_gm0_gm1_gn0_gn1 = descs[I2];
using AGridDesc_GK0_GM0_GM1_GK1 = decltype(a_grid_desc_gk0_gm0_gm1_gk1);
using BGridDesc_GK0_GN0_GN1_GK1 = decltype(b_grid_desc_gk0_gn0_gn1_gk1);
using CGridDesc_GM0_GM1_GN0_GN1 = decltype(c_grid_desc_gm0_gm1_gn0_gn1);
using AGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 0+: GK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 1+: GM0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2+: GM10
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 3+: GM11
Sequence<0, 0, 0, 0, 0, 0, 0>{}), // 4+: GK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 0-: GK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 1-: GM0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2-: GM10
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 3-: GM11
Sequence<0, 0, 0, 0, 0, 0, 0>{}))); // 4-: GK1
using BGridStepHacks = decltype(make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: GN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 2+: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 3+: GN11
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 4+: GK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: GN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 2-: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 3-: GN11
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}))); // 4-: GK1
using CGridStepHacks = decltype(make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GM10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 1+: BM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 2+: BM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 4+: BN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}), // 5+: GN1
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GM10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 1-: BM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 2-: BM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 4-: BN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}))); // 5-: GN1
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0>;
using BGridMoveSliceWindowStepHacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 2, 0, 0, 0, 0, 0>;
using GridwiseContraction =
GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN0_GN1<
BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
AGridDesc_GK0_GM0_GM1_GK1,
BGridDesc_GK0_GN0_GN1_GK1,
CGridDesc_GM0_GM1_GN0_GN1,
GM1PerBlockGM11,
GN1PerBlockGN11,
GK0PerBlock,
BM1PerThreadBM11,
BN1PerThreadBN11,
BK0PerThread,
BM10BN10ThreadClusterBM10Xs,
BM10BN10ThreadClusterBN10Xs,
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferSrcVectorTensorContiguousDimOrder,
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferSrcVectorTensorContiguousDimOrder,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks>;
if(get_block_1d_id() == 0 && get_thread_local_1d_id() == 0)
{
auto desc_tuple =
make_tuple(GridwiseContraction::MakeAGridDescriptor_GK0_GM0_GM10_GM11_GK1(
a_grid_desc_gk0_gm0_gm1_gk1),
GridwiseContraction::MakeBGridDescriptor_GK0_GN0_GN10_GN11_GK1(
b_grid_desc_gk0_gn0_gn1_gk1),
GridwiseContraction::MakeCGridDescriptor_GM10_BM0_BM1_GN10_BN0_BN1(
c_grid_desc_gm0_gm1_gn0_gn1),
GridwiseContraction::MakeCGridBlockCluster_BlockId_To_GM10_GN10(
c_grid_desc_gm0_gm1_gn0_gn1));
*static_cast<decltype(desc_tuple)*>(p_desc_tuple) = desc_tuple;
}
};
extern "C" __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_desc_tuple)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_n_c_hi_wi_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 28, 28));
constexpr auto wei_k_c_y_x_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 3, 3));
constexpr auto out_n_k_ho_wo_desc =
make_naive_tensor_descriptor_packed(make_tuple(256, 256, 28, 28));
constexpr auto descs =
transform_forward_convolution_into_contraction_v6r1_nchw_kcyx_nkhw_pad(wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
make_tuple(1, 1),
GN0,
GK1);
constexpr auto a_grid_desc_gk0_gm0_gm1_gk1 = descs[I0];
constexpr auto b_grid_desc_gk0_gn0_gn1_gk1 = descs[I1];
constexpr auto c_grid_desc_gm0_gm1_gn0_gn1 = descs[I2];
using AGridDesc_GK0_GM0_GM1_GK1 = decltype(a_grid_desc_gk0_gm0_gm1_gk1);
using BGridDesc_GK0_GN0_GN1_GK1 = decltype(b_grid_desc_gk0_gn0_gn1_gk1);
using CGridDesc_GM0_GM1_GN0_GN1 = decltype(c_grid_desc_gm0_gm1_gn0_gn1);
using AGridStepHacks =
decltype(make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 0+: GK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 1+: GM0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2+: GM10
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 3+: GM11
Sequence<0, 0, 0, 0, 0, 0, 0>{}), // 4+: GK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 0-: GK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 1-: GM0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2-: GM10
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 3-: GM11
Sequence<0, 0, 0, 0, 0, 0, 0>{}))); // 4-: GK1
using BGridStepHacks = decltype(make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: GN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 2+: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 3+: GN11
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 4+: GK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: GN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 2-: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 3-: GN11
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}))); // 4-: GK1
using CGridStepHacks = decltype(make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GM10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 1+: BM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 2+: BM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 4+: BN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}), // 5+: GN1
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GM10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 1-: BM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 2-: BM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 4-: BN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}))); // 5-: GN1
using AGridMoveSliceWindowStepHacks = Sequence<0, 0, 0, 0, 0, 0, 0>;
using BGridMoveSliceWindowStepHacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 2, 0, 0, 0, 0, 0>;
using GridwiseContraction =
GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN0_GN1<
BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
AGridDesc_GK0_GM0_GM1_GK1,
BGridDesc_GK0_GN0_GN1_GK1,
CGridDesc_GM0_GM1_GN0_GN1,
GM1PerBlockGM11,
GN1PerBlockGN11,
GK0PerBlock,
BM1PerThreadBM11,
BN1PerThreadBN11,
BK0PerThread,
BM10BN10ThreadClusterBM10Xs,
BM10BN10ThreadClusterBN10Xs,
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferSrcVectorTensorContiguousDimOrder,
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferSrcVectorTensorContiguousDimOrder,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks>;
using AGridDesc_GK0_GM0_GM10_GM11_GK1 =
decltype(GridwiseContraction::MakeAGridDescriptor_GK0_GM0_GM10_GM11_GK1(
a_grid_desc_gk0_gm0_gm1_gk1));
using BGridDesc_GK0_GN0_GN10_GN11_GK1 =
decltype(GridwiseContraction::MakeBGridDescriptor_GK0_GN0_GN10_GN11_GK1(
b_grid_desc_gk0_gn0_gn1_gk1));
using CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1 =
decltype(GridwiseContraction::MakeCGridDescriptor_GM10_BM0_BM1_GN10_BN0_BN1(
c_grid_desc_gm0_gm1_gn0_gn1));
using CGridBlockCluster_BlockId_To_GM10_GN10 =
decltype(GridwiseContraction::MakeCGridBlockCluster_BlockId_To_GM10_GN10(
c_grid_desc_gm0_gm1_gn0_gn1));
using DescTuple = decltype(make_tuple(AGridDesc_GK0_GM0_GM10_GM11_GK1{},
BGridDesc_GK0_GN0_GN10_GN11_GK1{},
CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1{},
CGridBlockCluster_BlockId_To_GM10_GN10{}));
const auto desc_tuple =
*reinterpret_cast<const DescTuple*>(cast_pointer_to_generic_address_space(p_desc_tuple));
const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = desc_tuple[I0];
const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = desc_tuple[I1];
const auto c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1 = desc_tuple[I2];
const auto c_grid_block_cluster_blockid_to_gm10_gn10 = desc_tuple[I3];
constexpr index_t shared_block_size =
GridwiseContraction::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseContraction::Run(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_block,
a_grid_desc_gk0_gm0_gm10_gm11_gk1,
b_grid_desc_gk0_gn0_gn10_gn11_gk1,
c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
c_grid_block_cluster_blockid_to_gm10_gn10,
integral_constant<bool, HasMainKBlockLoop>{},
integral_constant<bool, HasDoubleTailKBlockLoop>{});
};
device_operation/CMakeLists.txt deleted 100644 → 0
View file @ 245f7414
include_directories(BEFORE
include
${PROJECT_SOURCE_DIR}/host/host_tensor/include
${PROJECT_SOURCE_DIR}/device/include
${PROJECT_SOURCE_DIR}/device_operation/include
${PROJECT_SOURCE_DIR}/profiler/include
${PROJECT_SOURCE_DIR}/composable_kernel/include
${PROJECT_SOURCE_DIR}/composable_kernel/include/utility
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_operation
${PROJECT_SOURCE_DIR}/composable_kernel/include/problem_transform
${PROJECT_SOURCE_DIR}/external/rocm/include
)
# device_gemm_instance
set(DEVICE_GEMM_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp;
)
# device_gemm_bias_2d_instance
set(DEVICE_GEMM_BIAS_2D_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_nk_mn_instance.cpp;
)
# device_gemm_bias_relu_instance
set(DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instance.cpp;
)
# device_gemm_bias_relu_add_instance
set(DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instance.cpp;
)
set(DEVICE_BATCHED_GEMM_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp;
)
# device_conv2d_fwd_instance
set(DEVICE_CONV2D_FWD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp;
)
# device_conv1d_fwd_instance
set(DEVICE_CONV1D_FWD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instance.cpp;
)
# device_conv2d_fwd_bias_relu_instance
set(DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp;
)
# device_conv2d_fwd_bias_relu_add_instance
set(DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp;
)
# device_conv2d_fwd_bias_relu_atomic_add_instance
set(DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_bias_relu_atomic_add_nhwc_kyxc_nhwk_f16_instance.cpp;
)
# device_conv2d_bwd_data_instance
set(DEVICE_CONV2D_BWD_DATA_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp;
)
# device_reduce_instance
set(DEVICE_REDUCE_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_f16_f16_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_f16_f32_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_f32_f32_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_f32_f64_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_f64_f64_f64.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_threadwise_f16_f16_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_threadwise_f16_f32_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_threadwise_f32_f32_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_threadwise_f32_f64_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_threadwise_f64_f64_f64.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_second_call_f16_f16_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_second_call_f32_f32_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_second_call_f32_f32_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_second_call_f64_f64_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_blockwise_second_call_f64_f64_f64.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_atomic_add_f16_f32_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_atomic_add_f32_f32_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_atomic_add_f32_f64_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_partial_reduce_f16_f16_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_partial_reduce_f16_f32_f16.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_partial_reduce_f32_f32_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_partial_reduce_f32_f64_f32.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_reduce_instance_multiblock_partial_reduce_f64_f64_f64.cpp;
)
add_library(device_gemm_instance SHARED ${DEVICE_GEMM_INSTANCE_SOURCE})
add_library(device_gemm_bias_2d_instance SHARED ${DEVICE_GEMM_BIAS_2D_INSTANCE_SOURCE})
add_library(device_gemm_bias_relu_instance SHARED ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE})
add_library(device_gemm_bias_relu_add_instance SHARED ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE})
add_library(device_batched_gemm_instance SHARED ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE})
add_library(device_conv1d_fwd_instance SHARED ${DEVICE_CONV1D_FWD_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_instance SHARED ${DEVICE_CONV2D_FWD_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_atomic_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE})
add_library(device_conv2d_bwd_data_instance SHARED ${DEVICE_CONV2D_BWD_DATA_INSTANCE_SOURCE})
add_library(device_reduce_instance SHARED ${DEVICE_REDUCE_INSTANCE_SOURCE})
target_include_directories(device_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_gemm_bias_2d_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_gemm_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_gemm_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_batched_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv1d_fwd_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_atomic_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_bwd_data_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_reduce_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_compile_features(device_gemm_instance PUBLIC)
target_compile_features(device_gemm_bias_2d_instance PUBLIC)
target_compile_features(device_gemm_bias_relu_instance PUBLIC)
target_compile_features(device_gemm_bias_relu_add_instance PUBLIC)
target_compile_features(device_batched_gemm_instance PUBLIC)
target_compile_features(device_conv1d_fwd_instance PUBLIC)
target_compile_features(device_conv2d_fwd_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_add_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_atomic_add_instance PUBLIC)
target_compile_features(device_conv2d_bwd_data_instance PUBLIC)
target_compile_features(device_reduce_instance PUBLIC)
set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_gemm_bias_2d_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_gemm_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_gemm_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_batched_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv1d_fwd_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_atomic_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_bwd_data_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_reduce_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_gemm_instance LIBRARY DESTINATION lib)
install(TARGETS device_gemm_bias_2d_instance LIBRARY DESTINATION lib)
install(TARGETS device_gemm_bias_relu_instance LIBRARY DESTINATION lib)
install(TARGETS device_gemm_bias_relu_add_instance LIBRARY DESTINATION lib)
install(TARGETS device_batched_gemm_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv1d_fwd_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_add_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_atomic_add_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_bwd_data_instance LIBRARY DESTINATION lib)
install(TARGETS device_reduce_instance LIBRARY DESTINATION lib)
example/01_gemm/CMakeLists.txt 0 → 100644
View file @ 5d37d7bf
add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
example/02_gemm_alpha_beta/CMakeLists.txt 0 → 100644
View file @ 5d37d7bf
add_example_executable(example_gemm_xdl_alpha_beta gemm_xdl_alpha_beta.cpp)
example/03_gemm_bias_relu/CMakeLists.txt 0 → 100644
View file @ 5d37d7bf
add_example_executable(example_gemm_xdl_bias_relu gemm_xdl_bias_relu.cpp)
example/04_gemm_bias_relu_add/CMakeLists.txt 0 → 100644
View file @ 5d37d7bf
add_example_executable(example_gemm_xdl_bias_relu_add gemm_xdl_bias_relu_add.cpp)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment