Skip to content

GitLab

Commit 316d4acc authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin/develop' into aosewski/gemm_tile_loop

parents 9836e0ae 37a8c1f7
Hide whitespace changes
Inline Side-by-side
Showing with 1134 additions and 65 deletions
.pre-commit-config.yaml
View file @ 316d4acc
...@@ -3,7 +3,7 @@ repos: ...@@ -3,7 +3,7 @@ repos:
hooks: hooks:
- id: clang-format - id: clang-format
name: clang-format name: clang-format
entry: clang-format-10 -i --style=file entry: clang-format-12 -i --style=file
language: system language: system
types_or: [c++, inc] types_or: [c++, inc]
- id: copyright-year-checker - id: copyright-year-checker
......
CMakeLists.txt
View file @ 316d4acc
cmake_minimum_required(VERSION 3.14) cmake_minimum_required(VERSION 3.14)
set(version 1.1.0)
# Check support for CUDA/HIP in Cmake # Check support for CUDA/HIP in Cmake
project(composable_kernel) project(composable_kernel VERSION ${version})
list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake") list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
if (DTYPES) if (DTYPES)
add_definitions(-DDTYPES) add_definitions(-DDTYPES)
if (DTYPES MATCHES "int8") if (DTYPES MATCHES "int8")
add_definitions(-D__int8__) add_definitions(-DCK_ENABLE_INT8)
endif() set(CK_ENABLE_INT8 "ON")
if (DTYPES MATCHES "fp8") endif()
add_definitions(-D__fp8__) if (DTYPES MATCHES "fp8")
endif() add_definitions(-DCK_ENABLE_FP8)
if (DTYPES MATCHES "fp16") set(CK_ENABLE_FP8 "ON")
add_definitions(-D__fp16__) endif()
endif() if (DTYPES MATCHES "fp16")
if (DTYPES MATCHES "fp32") add_definitions(-DCK_ENABLE_FP16)
add_definitions(-D__fp32__) set(CK_ENABLE_FP16 "ON")
endif() endif()
if (DTYPES MATCHES "fp64") if (DTYPES MATCHES "fp32")
add_definitions(-D__fp64__) add_definitions(-DCK_ENABLE_FP32)
endif() set(CK_ENABLE_FP32 "ON")
if (DTYPES MATCHES "bf16") endif()
add_definitions(-D__bf16__) if (DTYPES MATCHES "fp64")
endif() add_definitions(-DCK_ENABLE_FP64)
message("DTYPES macro set to ${DTYPES}") set(CK_ENABLE_FP64 "ON")
endif()
if (DTYPES MATCHES "bf16")
add_definitions(-DCK_ENABLE_BF16)
set(CK_ENABLE_BF16 "ON")
endif()
message("DTYPES macro set to ${DTYPES}")
else() else()
add_definitions(-D__int8__ -D__fp8__ -D__fp16__ -D__fp32__ -D__fp64__ -D__bf16__) add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_ALL_DTYPES "ON")
endif() endif()
if(DL_KERNELS) if(DL_KERNELS)
add_definitions(-DDL_KERNELS) add_definitions(-DDL_KERNELS)
set(CK_ENABLE_DL_KERNELS "ON")
endif() endif()
if(INSTANCES_ONLY) if(INSTANCES_ONLY)
add_definitions(-DINSTANCES_ONLY) add_definitions(-DINSTANCES_ONLY)
set(CK_ENABLE_INSTANCES_ONLY "ON")
endif() endif()
# CK config file to record supported datatypes, etc.
configure_file("${PROJECT_SOURCE_DIR}/include/ck/config.h.in" "${PROJECT_BINARY_DIR}/include/ck/config.h")
# CK version file to record release version as well as git commit hash
find_package(Git REQUIRED)
execute_process(COMMAND "${GIT_EXECUTABLE}" rev-parse HEAD OUTPUT_VARIABLE COMMIT_ID OUTPUT_STRIP_TRAILING_WHITESPACE)
configure_file("${PROJECT_SOURCE_DIR}/include/ck/version.h.in" "${PROJECT_BINARY_DIR}/include/ck/version.h")
enable_testing() enable_testing()
set(ROCM_SYMLINK_LIBS OFF) set(ROCM_SYMLINK_LIBS OFF)
...@@ -50,8 +68,10 @@ include(ROCMInstallSymlinks) ...@@ -50,8 +68,10 @@ include(ROCMInstallSymlinks)
include(ROCMCreatePackage) include(ROCMCreatePackage)
include(CheckCXXCompilerFlag) include(CheckCXXCompilerFlag)
include(ROCMCheckTargetIds) include(ROCMCheckTargetIds)
rocm_setup_version(VERSION 0.2.0)
include(TargetFlags) include(TargetFlags)
rocm_setup_version(VERSION ${version})
list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip) list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip)
message("GPU_TARGETS= ${GPU_TARGETS}") message("GPU_TARGETS= ${GPU_TARGETS}")
...@@ -315,13 +335,14 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib) ...@@ -315,13 +335,14 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
set(CMAKE_ARCHIVE_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) set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
# set CK project include directories
include_directories(BEFORE include_directories(BEFORE
${PROJECT_BINARY_DIR}/include
${PROJECT_SOURCE_DIR}/include ${PROJECT_SOURCE_DIR}/include
${PROJECT_SOURCE_DIR}/library/include ${PROJECT_SOURCE_DIR}/library/include
${HIP_INCLUDE_DIRS} ${HIP_INCLUDE_DIRS}
) )
SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV") SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV")
if(BUILD_DEV) if(BUILD_DEV)
add_compile_options(-Werror) add_compile_options(-Werror)
...@@ -341,35 +362,35 @@ IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu ...@@ -341,35 +362,35 @@ IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu
file(READ "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}/CMakeLists.txt" cmake_instance) file(READ "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}/CMakeLists.txt" cmake_instance)
set(add_inst 0) set(add_inst 0)
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8") if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8")
#message("fp8 instance found!") #message("fp8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16") if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16")
#message("fp16 instance found!") #message("fp16 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32") if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32")
#message("fp32 instance found!") #message("fp32 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64") if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64")
#message("fp64 instance found!") #message("fp64 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16") if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16")
#message("bf16 instance found!") #message("bf16 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8") if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8")
#message("int8 instance found!") #message("int8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if(NOT "${cmake_instance}" MATCHES "DTYPES") if(NOT "${cmake_instance}" MATCHES "DTYPES")
#message("instance should be built for all types!") #message("instance should be built for all types!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if(add_inst EQUAL 1 OR NOT DEFINED DTYPES) if(add_inst EQUAL 1 OR NOT DEFINED DTYPES)
list(APPEND CK_DEVICE_INSTANCES device_${subdir_path}_instance) list(APPEND CK_DEVICE_INSTANCES device_${subdir_path}_instance)
endif() endif()
ENDIF() ENDIF()
ENDFOREACH() ENDFOREACH()
...@@ -409,7 +430,6 @@ endif() ...@@ -409,7 +430,6 @@ endif()
#Create an interface target for the include only files and call it "composablekernels" #Create an interface target for the include only files and call it "composablekernels"
include(CMakePackageConfigHelpers) include(CMakePackageConfigHelpers)
set(version 1.0.0)
write_basic_package_version_file( write_basic_package_version_file(
"${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake" "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
VERSION "${version}" VERSION "${version}"
...@@ -417,9 +437,9 @@ write_basic_package_version_file( ...@@ -417,9 +437,9 @@ write_basic_package_version_file(
) )
configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in
"${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake" "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
NO_CHECK_REQUIRED_COMPONENTS_MACRO NO_CHECK_REQUIRED_COMPONENTS_MACRO
) )
rocm_install(FILES rocm_install(FILES
...@@ -428,6 +448,13 @@ rocm_install(FILES ...@@ -428,6 +448,13 @@ rocm_install(FILES
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
) )
# Install CK version and configuration files
rocm_install(FILES
${PROJECT_BINARY_DIR}/include/ck/version.h
${PROJECT_BINARY_DIR}/include/ck/config.h
DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck/
)
set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE") set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE")
set(CPACK_RPM_PACKAGE_LICENSE "MIT") set(CPACK_RPM_PACKAGE_LICENSE "MIT")
......
Dockerfile
View file @ 316d4acc
...@@ -63,7 +63,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -63,7 +63,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
nano \ nano \
zlib1g-dev \ zlib1g-dev \
openssh-server \ openssh-server \
clang-format-10 \ clang-format-12 \
kmod && \ kmod && \
apt-get clean && \ apt-get clean && \
rm -rf /var/lib/apt/lists/* rm -rf /var/lib/apt/lists/*
......
Jenkinsfile
View file @ 316d4acc
...@@ -689,7 +689,7 @@ pipeline { ...@@ -689,7 +689,7 @@ pipeline {
-o -iname \'*.cpp.in\' \ -o -iname \'*.cpp.in\' \
-o -iname \'*.cl\' \ -o -iname \'*.cl\' \
| grep -v 'build/' \ | grep -v 'build/' \
| xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-10 -style=file {} | diff - {}\'" | xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-12 -style=file {} | diff - {}\'"
} }
steps{ steps{
buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true) buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
......
client_example/05_layernorm/layernorm2d.cpp
View file @ 316d4acc
...@@ -100,6 +100,10 @@ int main(int argc, char* argv[]) ...@@ -100,6 +100,10 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get())) if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N + std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
...@@ -153,6 +157,10 @@ int main(int argc, char* argv[]) ...@@ -153,6 +157,10 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get())) if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
} }
......
client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
View file @ 316d4acc
...@@ -191,6 +191,12 @@ int main(int argc, char* argv[]) ...@@ -191,6 +191,12 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get())) if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
} }
......
client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
View file @ 316d4acc
...@@ -187,6 +187,12 @@ int main(int argc, char* argv[]) ...@@ -187,6 +187,12 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get())) if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
} }
......
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ 316d4acc
...@@ -129,6 +129,10 @@ int main(int argc, char* argv[]) ...@@ -129,6 +129,10 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get())) if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte = std::size_t num_byte =
...@@ -184,6 +188,10 @@ int main(int argc, char* argv[]) ...@@ -184,6 +188,10 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get())) if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
} }
......
client_example/19_pool_fwd/CMakeLists.txt → client_example/19_pool/CMakeLists.txt
View file @ 316d4acc
add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp) add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations) target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_operations)
add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp) add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations) target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
\ No newline at end of file
add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_operations)
client_example/19_pool/avg_pool3d_bwd.cpp 0 → 100644
View file @ 316d4acc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
using DOutDataType = ck::half_t;
using DInDataType = ck::half_t;
using DOutLayout = ck::tensor_layout::convolution::NDHWC;
using DInLayout = ck::tensor_layout::convolution::NDHWC;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
std::size_t mMemSize_;
};
int main(int argc, char* argv[])
{
ck::index_t N = 2;
ck::index_t C = 32;
ck::index_t Z = 2;
ck::index_t Y = 2;
ck::index_t X = 2;
ck::index_t Di = 30;
ck::index_t Hi = 30;
ck::index_t Wi = 30;
ck::index_t window_stride_d = 2;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t window_dilation_d = 1;
ck::index_t window_dilation_h = 1;
ck::index_t window_dilation_w = 1;
ck::index_t in_left_pad_d = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_d = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
// Pool API only support the order of NCDHW
std::vector<ck::index_t> in_length = {N, C, Di, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Do, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
std::vector<ck::index_t> window_dilations{
window_dilation_d, window_dilation_h, window_dilation_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
std::size_t in_tensor_size = N * C * Di * Hi * Wi;
std::size_t out_tensor_size = N * C * Do * Ho * Wo;
// tensor layout = NDHWC
std::vector<ck::index_t> in_tensor_stride = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
using DeviceOp = ck::tensor_operation::device::
DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
// 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;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_length,
in_length,
out_tensor_stride,
in_tensor_stride,
window_spatial_lengths,
window_strides,
window_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
din_device_buf.SetZero();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes =
in_tensor_size * sizeof(DInDataType) + out_tensor_size * sizeof(DOutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_length,
in_length,
out_tensor_stride,
in_tensor_stride,
window_spatial_lengths,
window_strides,
window_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
din_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/19_pool/max_pool2d_bwd.cpp 0 → 100644
View file @ 316d4acc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using DOutDataType = ck::half_t;
using DInDataType = ck::half_t;
using IndexDataType = int32_t;
// We use pool3d to implement pool2d in this example
using InLayout = ck::tensor_layout::convolution::NDHWC;
using OutLayout = ck::tensor_layout::convolution::NDHWC;
constexpr ck::index_t InOutRank = 5;
constexpr ck::index_t WindowRank = 3;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
std::vector<ck::index_t>& window_lengths,
std::vector<ck::index_t>& output_lengths,
std::vector<ck::index_t>& input_stride,
std::vector<ck::index_t>& output_stride,
std::vector<ck::index_t>& indices_stride,
std::vector<ck::index_t>& window_strides,
std::vector<ck::index_t>& window_dilations,
std::vector<ck::index_t>& input_left_pads,
std::vector<ck::index_t>& input_right_pads,
std::vector<ck::index_t>& pooling_dims)
{
// NCHW to NCDHW
input_lengths.insert(input_lengths.begin() + 2, 1);
output_lengths.insert(output_lengths.begin() + 2, 1);
input_stride.insert(input_stride.begin() + 2, 0);
output_stride.insert(output_stride.begin() + 2, 0);
indices_stride.insert(indices_stride.begin() + 2, 0);
// YX to ZYX
window_lengths.insert(window_lengths.begin(), 1);
window_strides.insert(window_strides.begin(), 0);
window_dilations.insert(window_dilations.begin(), 0);
input_left_pads.insert(input_left_pads.begin(), 0);
input_right_pads.insert(input_right_pads.begin(), 0);
pooling_dims = {2, 3, 4};
}
int main(int argc, char* argv[])
{
ck::index_t N = 2;
ck::index_t C = 32;
ck::index_t Y = 2;
ck::index_t X = 2;
ck::index_t Hi = 30;
ck::index_t Wi = 30;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t window_dilation_h = 1;
ck::index_t window_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
// Pool API only support the order of NCHW
std::vector<ck::index_t> in_length = {N, C, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Y, X};
std::vector<ck::index_t> window_strides = {window_stride_h, window_stride_w};
std::vector<ck::index_t> window_dilations = {window_dilation_h, window_dilation_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_h, in_left_pad_w};
std::vector<ck::index_t> input_right_pads = {in_right_pad_h, in_right_pad_w};
std::vector<ck::index_t> pooling_dims = {2, 3};
std::size_t in_tensor_size = N * C * Hi * Wi;
std::size_t out_tensor_size = N * C * Ho * Wo;
// tensor layout = NHWC
std::vector<ck::index_t> in_tensor_stride = {C * Hi * Wi, 1, Wi * C, C};
std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
TransformPool2dparamToPool3d(in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
window_dilations,
input_left_pads,
input_right_pads,
pooling_dims);
SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
SimpleDeviceMem indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
// Generate index data from max pool forward
{
using MaxPoolFwdDeviceOp =
ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
InLayout,
OutLayout,
ck::ReduceTensorOp::MAX,
true>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
MaxPoolFwdDeviceOp>::GetInstances();
auto& op_ptr = op_ptrs[0];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
window_dilations,
input_left_pads,
input_right_pads,
pooling_dims);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
}
// Run MaxPool bwd
using MaxPoolBwdDeviceOp =
ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
MaxPoolBwdDeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_tensor_size,
in_tensor_size,
window_spatial_lengths,
window_strides,
window_dilations);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes = in_tensor_size * sizeof(DInDataType) +
out_tensor_size * sizeof(IndexDataType) +
out_tensor_size * sizeof(DOutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << "GB / s,"
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_tensor_size,
in_tensor_size,
window_spatial_lengths,
window_strides,
window_dilations);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/20_image_to_column/CMakeLists.txt 0 → 100644
View file @ 316d4acc
add_executable(client_image_to_column image_to_column.cpp)
target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)
client_example/20_image_to_column/image_to_column.cpp 0 → 100644
View file @ 316d4acc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::GNHWC;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 32; // batch size
static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 28; // input H
static constexpr ck::index_t Wi = 28; // input W
static constexpr ck::index_t Ho = 28; // output H
static constexpr ck::index_t Wo = 28; // output W
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, 2> in_spatial_lengths{Hi, Wi};
std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
// We have NHWGC in memory space (G is dummy)
// However, CK's API only accept length and stride with order of GNCHW
// Hence, we need to adjust the order of stride
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
std::array<ck::index_t, 2> out_strides{Y * X * C, 1};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * Y * X * C);
using DeviceOp = ck::tensor_operation::device::
DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>;
// 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;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
out.GetDeviceBuffer(),
N,
C,
in_spatial_lengths,
out_spatial_lengths,
wei_spatial_lengths,
in_strides,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
sizeof(OutDataType) * N * Ho * Wo * Y * X * C;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(avg_time < best_avg_time)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
out.GetDeviceBuffer(),
N,
C,
in_spatial_lengths,
out_spatial_lengths,
wei_spatial_lengths,
in_strides,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/21_grouped_gemm_bias/CMakeLists.txt 0 → 100644
View file @ 316d4acc
add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_operations)
client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp 0 → 100644
View file @ 316d4acc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include <random>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using D0DataType = F32;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = F32;
using ALayout = Row;
using BLayout = Row;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
int sum_of_m = 0;
Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
int group_count = Ms.size();
for(int i = 0; i < group_count; ++i)
{
Ns.push_back(768);
Ks.push_back(4608);
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
sum_of_m += Ms[i];
}
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<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, d0_dev_bufs, e_dev_bufs;
a_dev_bufs.reserve(group_count);
b_dev_bufs.reserve(group_count);
d0_dev_bufs.reserve(group_count);
e_dev_bufs.reserve(group_count);
std::vector<void*> p_e;
p_e.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; ++i)
{
a_dev_bufs.emplace_back(sizeof(ADataType) *
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
b_dev_bufs.emplace_back(sizeof(BDataType) *
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
d0_dev_bufs.emplace_back(sizeof(D0DataType) *
f_matrix_space_size(Ms[i], Ns[i], 0, D0Layout{}));
e_dev_bufs.emplace_back(sizeof(EDataType) *
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
grouped_gemm_kernel_args_.push_back(
{a_dev_bufs[i].GetDeviceBuffer(),
b_dev_bufs[i].GetDeviceBuffer(),
std::array<const void*, 1>{d0_dev_bufs[i].GetDeviceBuffer()},
e_dev_bufs[i].GetDeviceBuffer(),
Ms[i],
Ns[i],
Ks[i],
StrideAs[i],
StrideBs[i],
std::array<ck::index_t, 1>{0},
StrideEs[i]});
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// 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{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
std::vector<const void*> p_a = {}, p_b = {};
std::vector<std::array<const void*, 1>> p_ds = {};
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem grouped_gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
std::string op_name = op_ptr->GetTypeString();
hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetWorkSpacePointer(argument_ptr.get(),
grouped_gemm_workspace_dev.GetDeviceBuffer());
op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
grouped_gemm_kernel_args_dev.GetDeviceBuffer());
op_ptr->SetKBatch(argument_ptr.get(), 2);
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = 0, num_btype = 0;
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
{
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
sizeof(EDataType) * Ms[j] * Ns[j];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return 0;
}
client_example/CMakeLists.txt
View file @ 316d4acc
...@@ -3,31 +3,52 @@ project(ck_app) ...@@ -3,31 +3,52 @@ project(ck_app)
add_compile_options(-std=c++17) add_compile_options(-std=c++17)
if (DTYPES) if (DTYPES)
add_definitions(-DDTYPES) add_definitions(-DDTYPES)
if (DTYPES MATCHES "int8") if (DTYPES MATCHES "int8")
add_definitions(-D__int8__) add_definitions(-DCK_ENABLE_INT8)
if(NOT DEFINED ${CK_ENABLE_INT8})
set(CK_ENABLE_INT8 "ON")
endif() endif()
if (DTYPES MATCHES "fp8") endif()
add_definitions(-D__fp8__) if (DTYPES MATCHES "fp8")
add_definitions(-DCK_ENABLE_FP8)
if(NOT DEFINED ${CK_ENABLE_FP8})
set(CK_ENABLE_FP8 "ON")
endif() endif()
if (DTYPES MATCHES "fp16") endif()
add_definitions(-D__fp16__) if (DTYPES MATCHES "fp16")
add_definitions(-DCK_ENABLE_FP16)
if(NOT DEFINED ${CK_ENABLE_FP16})
set(CK_ENABLE_FP16 "ON")
endif() endif()
if (DTYPES MATCHES "fp32") endif()
add_definitions(-D__fp32__) if (DTYPES MATCHES "fp32")
add_definitions(-DCK_ENABLE_FP32)
if(NOT DEFINED ${CK_ENABLE_FP32})
set(CK_ENABLE_FP32 "ON")
endif() endif()
if (DTYPES MATCHES "fp64") endif()
add_definitions(-D__fp64__) if (DTYPES MATCHES "fp64")
add_definitions(-DCK_ENABLE_FP64)
if(NOT DEFINED ${CK_ENABLE_FP64})
set(CK_ENABLE_FP64 "ON")
endif() endif()
if (DTYPES MATCHES "bf16") endif()
add_definitions(-D__bf16__) if (DTYPES MATCHES "bf16")
add_definitions(-DCK_ENABLE_BF16)
if(NOT DEFINED ${CK_ENABLE_BF16})
set(CK_ENABLE_BF16 "ON")
endif() endif()
message("DTYPES macro set to ${DTYPES}") endif()
message("DTYPES macro set to ${DTYPES}")
else() else()
add_definitions(-D__int8__ -D__fp8__ -D__fp16__ -D__fp32__ -D__fp64__ -D__bf16__) add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
if(NOT DEFINED ${CK_ENABLE_ALL_DTYPES})
set(CK_ENABLE_ALL_DTYPES "ON")
endif()
endif() endif()
find_package(composable_kernel 1.0.0 COMPONENTS device_operations) find_package(composable_kernel COMPONENTS device_operations)
find_package(hip REQUIRED PATHS /opt/rocm) find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}") message(STATUS "Build with HIP ${hip_VERSION}")
......
cmake/DoxygenDoc.cmake
View file @ 316d4acc
...@@ -309,6 +309,8 @@ XML_OUTPUT ...@@ -309,6 +309,8 @@ XML_OUTPUT
XML_PROGRAMLISTING XML_PROGRAMLISTING
) )
set(WARN_AS_ERROR YES)
set(DOXYGEN_CONFIG_FILE "${CMAKE_CURRENT_BINARY_DIR}/doxygen/doxygen.conf" CACHE PATH "Path to generated doxygen configuration file") set(DOXYGEN_CONFIG_FILE "${CMAKE_CURRENT_BINARY_DIR}/doxygen/doxygen.conf" CACHE PATH "Path to generated doxygen configuration file")
function(add_doxygen_doc) function(add_doxygen_doc)
......
docs/Contributors_Guide.rst
View file @ 316d4acc
...@@ -2,7 +2,101 @@ ...@@ -2,7 +2,101 @@
Contributor's Guide Contributor's Guide
=================== ===================
Pull-request guidelines This chapter explains how to get started contributing to the Composable Kernel project and what are
======================= the contributing rules.
[TODO] Getting started
===============
#. **Documentation:** Before contributing to the library, familiarize yourself with the
`Composable Kernel User Guide <https://rocm.docs.amd.com/projects/composable_kernel/en/latest/>`_.
It provides insight into the core concepts, environment configuration, and steps to obtain or
build the library. You can also find some of this information in the
`README file <https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/README.md>`_
on the project's GitHub page.
#. **Additional reading:** We also recommend reading a `blog post
<https://community.amd.com/t5/instinct-accelerators/amd-composable-kernel-library-efficient-fused-kernels-for-ai/ba-p/553224>`_
from the AMD Community portal. It offers a deeper understanding of the library's objectives and
showcases its performance capabilities.
#. **General information:** For broader information about AMD products, consider exploring the
`AMD Developer Central portal <https://www.amd.com/en/developer.html>`_.
How do I contribute
===================
We deeply value contributions from our users. You can make an impact by reporting issues or
proposing code enhancements through pull requests.
Reporting issues
----------------
We use `Github issues <https://github.com/ROCmSoftwarePlatform/composable_kernel/issues>`_
to track public bugs and enhancement requests.
If you encounter an issue with the library, please check if the problem has already been
reported by searching existing issues on GitHub. If your issue seems unique, please submit a new
issue. All reported issues must include:
* A comprehensive description of the problem, including:
* What did you observe?
* Why do you think it is a bug (if it seems like one)?
* What did you expect to happen? What would indicate the resolution of the problem?
* Are there any known workarounds?
* Your configuration details, including:
* Which GPU are you using?
* Which OS version are you on?
* Which ROCm version are you using?
* Are you using a Docker image? If so, which one?
* Steps to reproduce the issue, including:
* What actions trigger the issue? What are the reproduction steps?
* If you build the library from scratch, what CMake command did you use?
* How frequently does this issue happen? Does it reproduce every time? Or is it a sporadic issue?
Before sumbitting any issue, ensure you have addressed all relevant questions from the checklist.
Creating Pull Requests
----------------------
You can submit `Pull Requests (PR) on GitHub
<https://github.com/ROCmSoftwarePlatform/composable_kernel/pulls>`_.
All contributors are required to develop their changes on a separate branch and then create a
pull requrest to merge their changes into the `develop` branch, which is the default
development branch in the Composable Kernel project. All external contributors must use their own
forks of the project to develop their changes.
When submitting a Pull Request you should:
* Describe the change providing information about the motivation for the change and a general
description of all code modifications.
* Verify and test the change:
* Run any relevant existing tests.
* Write new tests if added functionality is not covered by current tests.
* Ensure your changes align with the coding style defined in the ``.clang-format`` file located in
the project's root directory. We leverage `pre-commit` to run `clang-format` automatically. We
highly recommend contributors utilize this method to maintain consistent code formatting.
Instructions on setting up `pre-commit` can be found in the project's
`README file <https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/README.md>`_
* Link your PR to any related issues:
* If there is an issue that is resolved by your change, please provide a link to the issue in
the description of your pull request.
* For larger contributions, structure your change into a sequence of smaller, focused commits, each
addressing a particular aspect or fix.
Following the above guidelines ensures a seamless review process and faster assistance from our
end.
Thank you for your commitment to enhancing the Composable Kernel project! We look forward to collaborating with you.
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