Skip to content

GitLab

Commit 68886f7d authored by raman jana's avatar raman jana
Browse files

merging with latest develop branch

parents a9ee2960 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 1105 additions and 132 deletions
CMakeLists.txt
View file @ 68886f7d
...@@ -72,8 +72,9 @@ message(STATUS "Build with HIP ${HIP_VERSION}") ...@@ -72,8 +72,9 @@ message(STATUS "Build with HIP ${HIP_VERSION}")
rocm_create_package( rocm_create_package(
NAME CK-${CK_BACKEND} NAME composablekernel
DESCRIPTION "High Performance Composable Kernel for AMD GPUs" DESCRIPTION "High Performance Composable Kernel for AMD GPUs"
MAINTAINER "MIOpen Kernels Dev Team <dl.MIOpen@amd.com>"
LDCONFIG LDCONFIG
) )
...@@ -226,15 +227,12 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib) ...@@ -226,15 +227,12 @@ 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 ${CMAKE_CURRENT_BINARY_DIR}/bin)
configure_file("${PROJECT_SOURCE_DIR}/include/ck/hip_version.hpp.in" "${PROJECT_BINARY_DIR}/include/ck/hip_version.hpp")
include_directories(BEFORE include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include ${PROJECT_SOURCE_DIR}/include
${PROJECT_BINARY_DIR}/include ${PROJECT_BINARY_DIR}/include
${PROJECT_SOURCE_DIR}/library/include ${PROJECT_SOURCE_DIR}/library/include
) )
include(googletest)
SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV") SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV")
if(BUILD_DEV) if(BUILD_DEV)
...@@ -243,7 +241,31 @@ if(BUILD_DEV) ...@@ -243,7 +241,31 @@ if(BUILD_DEV)
endif() endif()
message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}") message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -C ${CMAKE_CFG_INTDIR})
add_subdirectory(library) add_subdirectory(library)
add_subdirectory(example) add_subdirectory(example)
add_subdirectory(test) add_subdirectory(test)
add_subdirectory(profiler) add_subdirectory(profiler)
#Create an interface target for the include only files and call it "composablekernels"
include(CMakePackageConfigHelpers)
set(version 1.0.0)
write_basic_package_version_file(
"${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
VERSION "${version}"
COMPATIBILITY AnyNewerVersion
)
configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in
"${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
NO_CHECK_REQUIRED_COMPONENTS_MACRO
)
install(FILES
"${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
"${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
Config.cmake.in 0 → 100644
View file @ 68886f7d
@PACKAGE_INIT@
set(_composable_kernel_supported_components device_operations host_tensor)
foreach(_comp ${composable_kernel_FIND_COMPONENTS})
if(NOT _comp IN_LIST _composable_kernel_supported_components)
set(composable_kernel_FOUND False)
set(composable_kernel_NOT_FOUND_MESSAGE "Unsupported component: ${_comp}")
endif()
include("${CMAKE_CURRENT_LIST_DIR}/composable_kernel${_comp}Targets.cmake")
endforeach()
Dockerfile
View file @ 68886f7d
...@@ -11,13 +11,7 @@ ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/ ...@@ -11,13 +11,7 @@ ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
RUN apt-get update RUN apt-get update
RUN apt-get install -y wget gnupg RUN apt-get install -y wget gnupg
RUN wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - RUN wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
RUN if ! [ -z $OSDB_BKC_VERSION ]; then \ RUN sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"
echo "Using BKC VERISION: $OSDB_BKC_VERSION";\
sh -c "echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-osdb-deb/ compute-rocm-dkms-no-npi-hipclang ${OSDB_BKC_VERSION} > /etc/apt/sources.list.d/rocm.list" ;\
cat /etc/apt/sources.list.d/rocm.list;\
else \
sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list" ;\
fi
RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add - RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
RUN sh -c "echo deb https://apt.kitware.com/ubuntu/ bionic main | tee -a /etc/apt/sources.list" RUN sh -c "echo deb https://apt.kitware.com/ubuntu/ bionic main | tee -a /etc/apt/sources.list"
...@@ -25,18 +19,15 @@ RUN sh -c "echo deb https://apt.kitware.com/ubuntu/ bionic main | tee -a /etc/ap ...@@ -25,18 +19,15 @@ RUN sh -c "echo deb https://apt.kitware.com/ubuntu/ bionic main | tee -a /etc/ap
# Install dependencies # Install dependencies
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
apt-utils \ apt-utils \
sshpass \
build-essential \ build-essential \
cmake-data=3.15.1-0kitware1 \ cmake-data=3.15.1-0kitware1 \
cmake=3.15.1-0kitware1 \ cmake=3.15.1-0kitware1 \
curl \ curl \
doxygen \
g++ \ g++ \
gdb \ gdb \
git \ git \
hip-rocclr \ hip-rocclr \
jq \ jq \
lcov \
libelf-dev \ libelf-dev \
libncurses5-dev \ libncurses5-dev \
libnuma-dev \ libnuma-dev \
...@@ -44,7 +35,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -44,7 +35,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
llvm-amdgpu \ llvm-amdgpu \
pkg-config \ pkg-config \
python \ python \
python3 \ python3.8 \
python-dev \ python-dev \
python3-dev \ python3-dev \
python-pip \ python-pip \
...@@ -62,8 +53,6 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -62,8 +53,6 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
apt-get clean && \ apt-get clean && \
rm -rf /var/lib/apt/lists/* rm -rf /var/lib/apt/lists/*
# RUN pip3 install --default-timeout=100000 -r requirements.txt
# Setup ubsan environment to printstacktrace # Setup ubsan environment to printstacktrace
RUN ln -s /usr/bin/llvm-symbolizer-3.8 /usr/local/bin/llvm-symbolizer RUN ln -s /usr/bin/llvm-symbolizer-3.8 /usr/local/bin/llvm-symbolizer
ENV UBSAN_OPTIONS=print_stacktrace=1 ENV UBSAN_OPTIONS=print_stacktrace=1
...@@ -83,6 +72,13 @@ ARG PREFIX=/opt/rocm ...@@ -83,6 +72,13 @@ ARG PREFIX=/opt/rocm
RUN cget install pfultz2/rocm-recipes RUN cget install pfultz2/rocm-recipes
# Install rbuild # Install rbuild
RUN pip3 install https://github.com/RadeonOpenCompute/rbuild/archive/6d78a0553babdaea8d2da5de15cbda7e869594b8.tar.gz RUN pip3 install https://github.com/RadeonOpenCompute/rbuild/archive/6d78a0553babdaea8d2da5de15cbda7e869594b8.tar.gz
# Install packages for processing the performance results
RUN pip3 install --upgrade pip
RUN pip3 install sqlalchemy
RUN pip3 install pymysql
RUN pip3 install pandas
RUN pip3 install setuptools-rust
RUN pip3 install sshtunnel
# Setup ubsan environment to printstacktrace # Setup ubsan environment to printstacktrace
ENV UBSAN_OPTIONS=print_stacktrace=1 ENV UBSAN_OPTIONS=print_stacktrace=1
...@@ -92,5 +88,3 @@ ADD rbuild.ini /rbuild.ini ...@@ -92,5 +88,3 @@ ADD rbuild.ini /rbuild.ini
ADD dev-requirements.txt dev-requirements.txt ADD dev-requirements.txt dev-requirements.txt
RUN rbuild prepare -s develop -d $PREFIX RUN rbuild prepare -s develop -d $PREFIX
RUN groupadd -f render RUN groupadd -f render
# RUN cget install -f min-requirements.txt
# RUN CXXFLAGS='-isystem $PREFIX/include' cget install -f ./mlir-requirements.txt
Jenkinsfile
View file @ 68886f7d
...@@ -212,16 +212,50 @@ def runCKProfiler(Map conf=[:]){ ...@@ -212,16 +212,50 @@ def runCKProfiler(Map conf=[:]){
{ {
cmake_build(conf) cmake_build(conf)
dir("script"){ dir("script"){
def perf_log = "perf_gemm_${gpu_arch}.log" //run gemm performance tests
def artifact = "profile_gemm_${gpu_arch}.txt" def gemm_log = "perf_gemm_${gpu_arch}.log"
sh "./profile_gemm.sh gemm 0 0 0 1 0 5 | tee ${perf_log} ||true" sh "rm -f ${gemm_log}"
sh "./profile_gemm.sh gemm 0 1 0 1 0 5 | tee -a ${perf_log} ||true" sh "echo Branch name: ${env.BRANCH_NAME} > ${gemm_log}"
sh "./profile_gemm.sh gemm 0 2 0 1 0 5 | tee -a ${perf_log} ||true" sh "echo Node name: ${NODE_NAME} >> ${gemm_log}"
sh "./profile_gemm.sh gemm 0 3 0 1 0 5 | tee -a ${perf_log} || true" sh "echo GPU_arch: ${gpu_arch} >> ${gemm_log}"
//results will be parsed, stored, and analyzed within the python script sh "hipcc --version | grep -e 'HIP version' >> ${gemm_log}"
//the script will return 0 if the performance criteria are met sh "/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}"
//or return 1 if the criteria are not met sh "./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}"
sh "python3 parse_perf_data.py ${perf_log} | tee ${artifact}" sh "./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 0 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 0 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 0 3 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a ${gemm_log}"
//results will be parsed, stored, and analyzed within the python script
//the script will return 0 if the performance criteria are met
//or return 1 if the criteria are not met
archiveArtifacts "${gemm_log}"
sh "python3 parse_perf_data.py ${gemm_log} "
//run resnet50 test
def resnet_log = "perf_resnet50_${gpu_arch}.log"
sh "rm -f ${resnet_log}"
sh "echo Branch name: ${env.BRANCH_NAME} > ${resnet_log}"
sh "echo Node name: ${NODE_NAME} >> ${resnet_log}"
sh "echo GPU_arch: ${gpu_arch} >> ${resnet_log}"
sh "hipcc --version | grep -e 'HIP version' >> ${resnet_log}"
sh "/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet_log}"
//first run tests with N=256
sh "./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a ${resnet_log}"
//then run with N=4
sh "./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a ${resnet_log}"
archiveArtifacts "${resnet_log}"
//the script will put the results from N=256 and N=4 runs into separate tables
sh "python3 parse_perf_data.py ${resnet_log} "
} }
} }
} }
...@@ -246,7 +280,6 @@ def runPerfTest(Map conf=[:]){ ...@@ -246,7 +280,6 @@ def runPerfTest(Map conf=[:]){
} }
} }
pipeline { pipeline {
agent none agent none
options { options {
...@@ -280,19 +313,19 @@ pipeline { ...@@ -280,19 +313,19 @@ pipeline {
// buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release') // buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
// } // }
//} //}
stage('Build Profiler: Debug, gfx908') //stage('Build Profiler: Debug, gfx908')
{ //{
agent { label rocmnode("nogpu")} // agent { label rocmnode("nogpu")}
environment{ // environment{
setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " -DBUILD_DEV=On """ // setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " -DBUILD_DEV=On """
} // }
steps{ // steps{
// until we stabilize debug build due to compiler crashes // // until we stabilize debug build due to compiler crashes
catchError(buildResult: 'SUCCESS', stageResult: 'FAILURE') { // catchError(buildResult: 'SUCCESS', stageResult: 'FAILURE') {
buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Debug') // buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Debug')
} // }
} // }
} //}
stage('Clang Format') { stage('Clang Format') {
agent{ label rocmnode("nogpu") } agent{ label rocmnode("nogpu") }
environment{ environment{
...@@ -312,7 +345,7 @@ pipeline { ...@@ -312,7 +345,7 @@ pipeline {
} }
} }
} }
stage("Tests") stage("Tests")
{ {
parallel parallel
{ {
...@@ -320,7 +353,7 @@ pipeline { ...@@ -320,7 +353,7 @@ pipeline {
{ {
agent{ label rocmnode("gfx908")} agent{ label rocmnode("gfx908")}
environment{ environment{
setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " -DBUILD_DEV=On """ setup_args = """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx900 --offload-arch=gfx906 --offload-arch=gfx908 --offload-arch=gfx90a -O3 " -DBUILD_DEV=On """
} }
steps{ steps{
buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "check", no_reboot:true, build_type: 'Release') buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "check", no_reboot:true, build_type: 'Release')
...@@ -341,6 +374,23 @@ pipeline { ...@@ -341,6 +374,23 @@ pipeline {
} }
} }
stage("Client App")
{
parallel
{
stage("Run Client App")
{
agent{ label rocmnode("gfx908")}
environment{
setup_args = """ -D -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """
execute_args = """ cd ../test/client_app && rm -rf build && mkdir build && cd build && cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" .. && make """
}
steps{
buildHipClangJobAndReboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
}
}
}
}
stage("Performance Tests") stage("Performance Tests")
{ {
parallel parallel
...@@ -350,15 +400,20 @@ pipeline { ...@@ -350,15 +400,20 @@ pipeline {
agent{ label rocmnode("gfx908")} agent{ label rocmnode("gfx908")}
environment{ environment{
setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " -DBUILD_DEV=On """ setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " -DBUILD_DEV=On """
} dbuser = "${dbuser}"
dbpassword = "${dbpassword}"
dbsship = "${dbsship}"
dbsshport = "${dbsshport}"
dbsshuser = "${dbsshuser}"
dbsshpassword = "${dbsshpassword}"
}
steps{ steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release') runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
} }
} }
} }
} }
// enable after the cmake file supports packaging // enable after the cmake file supports packaging
// stage("Packages") { // stage("Packages") {
// when { // when {
......
README.md
View file @ 68886f7d
...@@ -43,3 +43,13 @@ Instructions for running each individual examples are under ```example/``` ...@@ -43,3 +43,13 @@ Instructions for running each individual examples are under ```example/```
make -j ckProfiler make -j ckProfiler
``` ```
Instructions for running ckProfiler are under ```profiler/``` Instructions for running ckProfiler are under ```profiler/```
## Caveat
### Kernel Timing and Verification
CK's own kernel timer will warn up kernel once, and then run it multiple times
to get average kernel time. For some kernels that use atomic add, this will cause
output buffer to be accumulated multiple times, causing verfication failure.
To work around it, do not use CK's own timer and do verification at the same time.
CK's own timer and verification in each example and ckProfiler can be enabled or
disabled from command line.
cmake/googletest.cmake
View file @ 68886f7d
...@@ -19,6 +19,7 @@ list(APPEND GTEST_CMAKE_CXX_FLAGS ...@@ -19,6 +19,7 @@ list(APPEND GTEST_CMAKE_CXX_FLAGS
-Wno-zero-as-null-pointer-constant -Wno-zero-as-null-pointer-constant
-Wno-unused-member-function -Wno-unused-member-function
-Wno-comma -Wno-comma
-Wno-old-style-cast
) )
message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}") message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
...@@ -35,4 +36,4 @@ FetchContent_MakeAvailable(googletest) ...@@ -35,4 +36,4 @@ FetchContent_MakeAvailable(googletest)
target_compile_options(gtest PRIVATE ${GTEST_CMAKE_CXX_FLAGS}) target_compile_options(gtest PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
target_compile_options(gtest_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS}) target_compile_options(gtest_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
target_compile_options(gmock PRIVATE ${GTEST_CMAKE_CXX_FLAGS}) target_compile_options(gmock PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
target_compile_options(gmock_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
example/01_gemm/CMakeLists.txt
View file @ 68886f7d
add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp) 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_bf16 gemm_xdl_bf16.cpp)
add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp) add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
example/01_gemm/gemm_dl_fp16.cpp 0 → 100644
View file @ 68886f7d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
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 ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = Col;
using BLayout = Row;
using CLayout = Row;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::
// ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 1)
{
// do nothing
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(1);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
bool pass = true;
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
pass = ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
}
return pass ? 0 : 1;
}
example/01_gemm/gemm_dl_fp32.cpp 0 → 100644
View file @ 68886f7d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
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 ADataType = float;
using BDataType = float;
using CDataType = float;
using AccDataType = float;
using ALayout = Col;
using BLayout = Row;
using CLayout = Row;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::
// ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 1)
{
// do nothing
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(1);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
bool pass = true;
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
pass = ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
}
return pass ? 0 : 1;
}
example/01_gemm/gemm_dl_int8.cpp 0 → 100644
View file @ 68886f7d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using ALayout = Col;
using BLayout = Row;
using CLayout = Row;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::
// #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// #########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// #########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// #########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< int8_t, int8_t, int8_t, int32_t, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 1)
{
// do nothing
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(1);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
bool pass = true;
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
pass = ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
}
return pass ? 0 : 1;
}
example/01_gemm/gemm_xdl_bf16.cpp
View file @ 68886f7d
...@@ -84,13 +84,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -84,13 +84,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<float, float, float, PassThrough, PassThrough, PassThrough>; ReferenceGemm<float, float, float, float, PassThrough, PassThrough, PassThrough>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 3840;
...@@ -105,13 +105,13 @@ int main(int argc, char* argv[]) ...@@ -105,13 +105,13 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 10) else if(argc == 10)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); M = std::stoi(argv[4]);
N = std::stoi(argv[5]); N = std::stoi(argv[5]);
...@@ -125,7 +125,7 @@ int main(int argc, char* argv[]) ...@@ -125,7 +125,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n"); printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0); exit(0);
} }
...@@ -193,12 +193,12 @@ int main(int argc, char* argv[]) ...@@ -193,12 +193,12 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = std::size_t num_btype =
...@@ -232,7 +232,7 @@ int main(int argc, char* argv[]) ...@@ -232,7 +232,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0; return 0;
......
example/01_gemm/gemm_xdl_fp16.cpp
View file @ 68886f7d
...@@ -62,13 +62,13 @@ using DeviceGemmInstance_WaveletModel = ck::tensor_operation::device::DeviceGemm ...@@ -62,13 +62,13 @@ using DeviceGemmInstance_WaveletModel = ck::tensor_operation::device::DeviceGemm
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 3840;
...@@ -83,13 +83,13 @@ int main(int argc, char* argv[]) ...@@ -83,13 +83,13 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 10) else if(argc == 10)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); M = std::stoi(argv[4]);
N = std::stoi(argv[5]); N = std::stoi(argv[5]);
...@@ -103,7 +103,7 @@ int main(int argc, char* argv[]) ...@@ -103,7 +103,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n"); printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0); exit(0);
} }
...@@ -178,12 +178,12 @@ int main(int argc, char* argv[]) ...@@ -178,12 +178,12 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = std::size_t num_btype =
...@@ -208,7 +208,7 @@ int main(int argc, char* argv[]) ...@@ -208,7 +208,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0; return 0;
......
example/01_gemm/gemm_xdl_fp64.cpp 0 → 100644
View file @ 68886f7d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F64 = double;
using ADataType = double;
using BDataType = double;
using CDataType = double;
using AccDataType = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if 0
< F64, F64, F64, F64, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 32, 4, 1, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 7, 1>;
#else
< F64, F64, F64, F64, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 2, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1>;
#endif
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
template <typename DataType>
std::ostream& show_2d_matrix(std::ostream& os, Tensor<DataType>& matrix)
{
os << "[" << std::endl;
for(int x = 0; x < matrix.mDesc.GetLengths()[0]; x++)
{
os << "[";
for(int y = 0; y < matrix.mDesc.GetLengths()[1]; y++)
{
os << std::setw(4) << static_cast<float>(matrix(x, y));
}
os << "]" << std::endl;
}
os << "]";
return os;
}
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "data type: " << typeid(ADataType{}).name() << std::endl;
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
#if 0
{
show_2d_matrix(std::cout << "a : ", a_m_k) << std::endl;
show_2d_matrix(std::cout << "b: ", b_k_n) << std::endl;
show_2d_matrix(std::cout << "c_device: ", c_m_n_device_result) << std::endl;
show_2d_matrix(std::cout << "c_host :", c_m_n_host_result) << std::endl;
}
#endif
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}
example/01_gemm/gemm_xdl_int8.cpp
View file @ 68886f7d
...@@ -78,14 +78,19 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -78,14 +78,19 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
16>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock 16>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>; BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 3840;
...@@ -100,13 +105,13 @@ int main(int argc, char* argv[]) ...@@ -100,13 +105,13 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 10) else if(argc == 10)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); M = std::stoi(argv[4]);
N = std::stoi(argv[5]); N = std::stoi(argv[5]);
...@@ -120,7 +125,7 @@ int main(int argc, char* argv[]) ...@@ -120,7 +125,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n"); printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0); exit(0);
} }
...@@ -189,12 +194,12 @@ int main(int argc, char* argv[]) ...@@ -189,12 +194,12 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = std::size_t num_btype =
...@@ -219,7 +224,7 @@ int main(int argc, char* argv[]) ...@@ -219,7 +224,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0; return 0;
......
example/02_gemm_alpha_beta/gemm_xdl_alpha_beta.cpp
View file @ 68886f7d
...@@ -86,9 +86,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<AD ...@@ -86,9 +86,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<AD
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 3840;
...@@ -106,13 +106,13 @@ int main(int argc, char* argv[]) ...@@ -106,13 +106,13 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 6) else if(argc == 6)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]); alpha = std::stof(argv[4]);
beta = std::stof(argv[5]); beta = std::stof(argv[5]);
...@@ -121,7 +121,7 @@ int main(int argc, char* argv[]) ...@@ -121,7 +121,7 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); M = std::stoi(argv[4]);
N = std::stoi(argv[5]); N = std::stoi(argv[5]);
...@@ -138,7 +138,7 @@ int main(int argc, char* argv[]) ...@@ -138,7 +138,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n"); printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, alpha, beta\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, alpha, beta\n");
exit(0); exit(0);
} }
...@@ -216,7 +216,7 @@ int main(int argc, char* argv[]) ...@@ -216,7 +216,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem"); "not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = std::size_t num_btype =
...@@ -246,6 +246,8 @@ int main(int argc, char* argv[]) ...@@ -246,6 +246,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0;
} }
example/03_gemm_bias_relu/gemm_xdl_bias_relu.cpp
View file @ 68886f7d
...@@ -83,9 +83,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActiv ...@@ -83,9 +83,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActiv
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 3840;
...@@ -100,13 +100,13 @@ int main(int argc, char* argv[]) ...@@ -100,13 +100,13 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 10) else if(argc == 10)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); M = std::stoi(argv[4]);
N = std::stoi(argv[5]); N = std::stoi(argv[5]);
...@@ -120,7 +120,7 @@ int main(int argc, char* argv[]) ...@@ -120,7 +120,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n"); printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0); exit(0);
} }
...@@ -206,7 +206,7 @@ int main(int argc, char* argv[]) ...@@ -206,7 +206,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem"); "not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
...@@ -232,6 +232,8 @@ int main(int argc, char* argv[]) ...@@ -232,6 +232,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0;
} }
example/04_gemm_bias_relu_add/gemm_xdl_bias_relu_add.cpp
View file @ 68886f7d
...@@ -83,9 +83,9 @@ using ReferenceGemmInstance = ...@@ -83,9 +83,9 @@ using ReferenceGemmInstance =
CElementOp>; CElementOp>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 3840;
...@@ -101,13 +101,13 @@ int main(int argc, char* argv[]) ...@@ -101,13 +101,13 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 11) else if(argc == 11)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); M = std::stoi(argv[4]);
N = std::stoi(argv[5]); N = std::stoi(argv[5]);
...@@ -122,7 +122,7 @@ int main(int argc, char* argv[]) ...@@ -122,7 +122,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n"); printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\n"); printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\n");
exit(0); exit(0);
} }
...@@ -218,7 +218,7 @@ int main(int argc, char* argv[]) ...@@ -218,7 +218,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem"); "not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
...@@ -250,6 +250,8 @@ int main(int argc, char* argv[]) ...@@ -250,6 +250,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0;
} }
example/06_conv2d_fwd_bias_relu/conv2d_fwd_xdl_bias_relu.cpp
View file @ 68886f7d
...@@ -93,7 +93,7 @@ void PrintUseMsg() ...@@ -93,7 +93,7 @@ void PrintUseMsg()
{ {
std::cout << "arg1: verification (0=no, 1=yes)\n" std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n" << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n" << "arg3: time kernel (0=n0, 1=yes)\n"
<< "Following arguments:\n" << "Following arguments:\n"
<< " N, K, C, \n" << " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n" << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
...@@ -165,9 +165,9 @@ int main(int argc, char* argv[]) ...@@ -165,9 +165,9 @@ int main(int argc, char* argv[])
{ {
using namespace ck::utils::conv; using namespace ck::utils::conv;
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
const int num_dim_spatial = 2; const int num_dim_spatial = 2;
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
...@@ -176,7 +176,7 @@ int main(int argc, char* argv[]) ...@@ -176,7 +176,7 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
if(argc >= 5) if(argc >= 5)
...@@ -269,7 +269,7 @@ int main(int argc, char* argv[]) ...@@ -269,7 +269,7 @@ int main(int argc, char* argv[])
"not support this problem"); "not support this problem");
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops( std::size_t flop = get_flops(
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths); params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
...@@ -305,7 +305,8 @@ int main(int argc, char* argv[]) ...@@ -305,7 +305,8 @@ int main(int argc, char* argv[])
OutElementOp{}); OutElementOp{});
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data()); out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err( return ck::utils::check_err(device_output.mData, host_output.mData) ? 0 : 1;
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
} }
return 0;
} }
example/07_conv2d_fwd_bias_relu_add/CMakeLists.txt
View file @ 68886f7d
add_example_executable(example_conv2d_fwd_xdl_bias_relu_add conv2d_fwd_xdl_bias_relu_add.cpp) # FIXME: should fix validation failure
add_example_executable_no_testing(example_conv2d_fwd_xdl_bias_relu_add conv2d_fwd_xdl_bias_relu_add.cpp)
target_link_libraries(example_conv2d_fwd_xdl_bias_relu_add PRIVATE conv_util) target_link_libraries(example_conv2d_fwd_xdl_bias_relu_add PRIVATE conv_util)
example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp
View file @ 68886f7d
...@@ -90,7 +90,7 @@ void PrintUseMsg() ...@@ -90,7 +90,7 @@ void PrintUseMsg()
{ {
std::cout << "arg1: verification (0=no, 1=yes)\n" std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n" << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n" << "arg3: time kernel (0=n0, 1=yes)\n"
<< "Following arguments:\n" << "Following arguments:\n"
<< " N, K, C, \n" << " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n" << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
...@@ -162,9 +162,9 @@ int main(int argc, char* argv[]) ...@@ -162,9 +162,9 @@ int main(int argc, char* argv[])
{ {
using namespace ck::utils::conv; using namespace ck::utils::conv;
bool do_verification = 0; bool do_verification = true;
int init_method = 0; int init_method = 1;
int nrepeat = 5; bool time_kernel = false;
const int num_dim_spatial = 2; const int num_dim_spatial = 2;
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
...@@ -173,7 +173,7 @@ int main(int argc, char* argv[]) ...@@ -173,7 +173,7 @@ int main(int argc, char* argv[])
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
if(argc >= 5) if(argc >= 5)
...@@ -224,10 +224,10 @@ int main(int argc, char* argv[]) ...@@ -224,10 +224,10 @@ int main(int argc, char* argv[])
{ {
case 0: break; case 0: break;
case 1: case 1:
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
residual.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); residual.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
break; break;
default: default:
input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
...@@ -280,7 +280,7 @@ int main(int argc, char* argv[]) ...@@ -280,7 +280,7 @@ int main(int argc, char* argv[])
"not support this problem"); "not support this problem");
} }
float ave_time = invoker.Run(argument, nrepeat); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops( std::size_t flop = get_flops(
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths); params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
...@@ -320,7 +320,8 @@ int main(int argc, char* argv[]) ...@@ -320,7 +320,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data()); out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err( return ck::utils::check_err(device_output.mData, host_output.mData) ? 0 : 1;
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
} }
return 0;
} }
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