Skip to content

GitLab

Commit aa5859e4 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into wavelet_model

parents 9bd6cc0e 5ee30459
Show whitespace changes
Inline Side-by-side
Showing with 1129 additions and 203 deletions
CMakeLists.txt
View file @ aa5859e4
......@@ -8,7 +8,7 @@ list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
enable_testing()
set(ROCM_SYMLINK_LIBS OFF)
find_package(ROCM 0.8 REQUIRED PATHS /opt/rocm)
find_package(ROCM REQUIRED PATHS /opt/rocm)
include(ROCMInstallTargets)
include(ROCMPackageConfigHelpers)
......@@ -71,13 +71,6 @@ if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
endif()
message(STATUS "Build with HIP ${HIP_VERSION}")
rocm_create_package(
NAME composablekernel
DESCRIPTION "High Performance Composable Kernel for AMD GPUs"
MAINTAINER "MIOpen Kernels Dev Team <dl.MIOpen@amd.com>"
LDCONFIG
)
## tidy
include(EnableCompilerWarnings)
set(CK_TIDY_ERRORS ERRORS * -readability-inconsistent-declaration-parameter-name)
......
Dockerfile
View file @ aa5859e4
......@@ -2,6 +2,7 @@ FROM ubuntu:18.04
ARG ROCMVERSION=5.1
ARG OSDB_BKC_VERSION
ARG compiler_version
RUN set -xe
......@@ -15,7 +16,6 @@ RUN sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.l
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"
# ADD requirements.txt requirements.txt
# Install dependencies
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
apt-utils \
......@@ -23,8 +23,6 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
cmake-data=3.15.1-0kitware1 \
cmake=3.15.1-0kitware1 \
curl \
g++ \
gdb \
git \
hip-rocclr \
jq \
......@@ -61,17 +59,7 @@ ENV UBSAN_OPTIONS=print_stacktrace=1
RUN wget https://github.com/Yelp/dumb-init/releases/download/v1.2.0/dumb-init_1.2.0_amd64.deb
RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
# Install cget
RUN pip install cget
# Install rclone
RUN pip install https://github.com/pfultz2/rclone/archive/master.tar.gz
ARG PREFIX=/opt/rocm
# Install dependencies
RUN cget install pfultz2/rocm-recipes
# Install rbuild
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
......@@ -84,12 +72,26 @@ ENV UBSAN_OPTIONS=print_stacktrace=1
ENV LC_ALL=C.UTF-8
ENV LANG=C.UTF-8
ADD rbuild.ini /rbuild.ini
ADD dev-requirements.txt dev-requirements.txt
RUN rbuild prepare -s develop -d $PREFIX
RUN groupadd -f render
# Install the new rocm-cmake version
RUN git clone -b master https://github.com/RadeonOpenCompute/rocm-cmake.git && \
cd rocm-cmake && mkdir build && cd build && \
cmake .. && cmake --build . && cmake --build . --target install
WORKDIR /
ENV compiler_version=$compiler_version
RUN sh -c "echo compiler version = '$compiler_version'"
RUN --mount=type=ssh if [ "$compiler_version" != "release" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
make -j 8 ; \
else echo "using the release compiler"; \
fi
#ENV HIP_CLANG_PATH='/llvm-project/build/bin'
#RUN sh -c "echo HIP_CLANG_PATH = '$HIP_CLANG_PATH'"
Jenkinsfile
View file @ aa5859e4
......@@ -11,6 +11,96 @@ def show_node_info() {
"""
}
def runShell(String command){
def responseCode = sh returnStatus: true, script: "${command} > tmp.txt"
def output = readFile(file: "tmp.txt")
echo "tmp.txt contents: $output"
return (output != "")
}
def getDockerImageName(){
def img = "${env.MIOPEN_IMAGE_URL}:composable_kernels_${params.COMPILER_VERSION}"
return img
}
def getDockerImage(Map conf=[:]){
env.DOCKER_BUILDKIT=1
def prefixpath = conf.get("prefixpath", "/opt/rocm") // prefix:/opt/rocm
def gpu_arch = conf.get("gpu_arch", "gfx908") // prebuilt dockers should have all the architectures enabled so one image can be used for all stages
def no_cache = conf.get("no_cache", false)
def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' "
if(env.CCACHE_HOST)
{
def check_host = sh(script:"""(printf "PING\r\n";) | nc -N ${env.CCACHE_HOST} 6379 """, returnStdout: true).trim()
if(check_host == "+PONG")
{
echo "FOUND CCACHE SERVER: ${CCACHE_HOST}"
}
else
{
echo "CCACHE SERVER: ${CCACHE_HOST} NOT FOUND, got ${check_host} response"
}
dockerArgs = dockerArgs + " --build-arg CCACHE_SECONDARY_STORAGE='redis://${env.CCACHE_HOST}' --build-arg COMPILER_LAUNCHER='ccache' "
env.CCACHE_DIR = """/tmp/ccache_store"""
env.CCACHE_SECONDARY_STORAGE="""redis://${env.CCACHE_HOST}"""
}
if(no_cache)
{
dockerArgs = dockerArgs + " --no-cache "
}
echo "Docker Args: ${dockerArgs}"
def image = getDockerImageName()
//Check if image exists
def retimage
try
{
echo "Pulling down image: ${image}"
retimage = docker.image("${image}")
retimage.pull()
}
catch(Exception ex)
{
error "Unable to locate image: ${image}"
}
return [retimage, image]
}
def buildDocker(install_prefix){
show_node_info()
env.DOCKER_BUILDKIT=1
checkout scm
def image_name = getDockerImageName()
echo "Building Docker for ${image_name}"
def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg compiler_version='${params.COMPILER_VERSION}' "
if(env.CCACHE_HOST)
{
def check_host = sh(script:"""(printf "PING\\r\\n";) | nc -N ${env.CCACHE_HOST} 6379 """, returnStdout: true).trim()
if(check_host == "+PONG")
{
echo "FOUND CCACHE SERVER: ${CCACHE_HOST}"
}
else
{
echo "CCACHE SERVER: ${CCACHE_HOST} NOT FOUND, got ${check_host} response"
}
dockerArgs = dockerArgs + " --build-arg CCACHE_SECONDARY_STORAGE='redis://${env.CCACHE_HOST}' --build-arg COMPILER_LAUNCHER='ccache' "
env.CCACHE_DIR = """/tmp/ccache_store"""
env.CCACHE_SECONDARY_STORAGE="""redis://${env.CCACHE_HOST}"""
}
echo "Build Args: ${dockerArgs}"
try{
echo "Checking for image: ${image_name}"
sh "docker manifest inspect --insecure ${image_name}"
echo "Image: ${image_name} found!! Skipping building image"
}
catch(Exception ex){
echo "Unable to locate image: ${image_name}. Building image now"
retimage = docker.build("${image_name}", dockerArgs + ' .')
retimage.push()
}
}
def cmake_build(Map conf=[:]){
def compiler = conf.get("compiler","/opt/rocm/bin/hipcc")
......@@ -60,7 +150,7 @@ def cmake_build(Map conf=[:]){
"""
def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
// reduce parallelism when compiling, clang uses too much memory
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(( \$(nproc) / 1 )) ${config_targets}")
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(( \$(nproc) / 2 )) ${config_targets}")
def execute_cmd = conf.get("execute_cmd", "")
def cmd = conf.get("cmd", """
......@@ -85,7 +175,7 @@ def buildHipClangJob(Map conf=[:]){
env.HSA_ENABLE_SDMA=0
checkout scm
def image = "composable_kernels"
def image = "composable_kernels_${params.COMPILER_VERSION}"
def prefixpath = conf.get("prefixpath", "/opt/rocm")
def gpu_arch = conf.get("gpu_arch", "gfx908")
......@@ -93,22 +183,31 @@ def buildHipClangJob(Map conf=[:]){
// def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1"
dockerOpts = dockerOpts + " --env HSA_XNACK=1 --env GPU_ARCH='${gpu_arch}' "
}
//def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg GPU_ARCH='${gpu_arch}' --build-arg compiler_version='${params.COMPILER_VERSION}' "
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' "
if (params.COMPILER_VERSION != "release"){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg GPU_ARCH='${gpu_arch}' "
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
if (params.USE_DOCKERFILE){
try {
retimage = docker.build("${image}", dockerArgs + '.')
//retimage = docker.build("${image}", dockerArgs + '.')
(retimage, image) = getDockerImage(conf)
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES')
{
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
throw new Exception ("GPU not found")
}
else{
echo "GPU is OK"
}
}
}
}
......@@ -117,27 +216,23 @@ def buildHipClangJob(Map conf=[:]){
throw e
}
catch(Exception ex) {
retimage = docker.build("${image}", dockerArgs + "--no-cache .")
retimage = docker.build("${image}", dockerArgs + " --no-cache .")
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES')
{
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo |tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
throw new Exception ("GPU not found")
}
else{
echo "GPU is OK"
}
}
}
else{
timeout(time: 3, unit: 'HOURS'){
retimage = docker.image('compute-artifactory.amd.com:5000/rocm-plus-docker/framework/compute-rocm-dkms-no-npi-hipclang:9110_ubuntu18.04_py3.6_pytorch_rocm5.0_internal_testing_7ff5b54').pull()
image="b56f8ac0d6ea"
sh "docker images"
}
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 5, unit: 'HOURS')
{
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
cmake_build(conf)
}
}
......@@ -149,10 +244,6 @@ def reboot(){
build job: 'reboot-slaves', propagate: false , parameters: [string(name: 'server', value: "${env.NODE_NAME}"),]
}
def buildHipClangJobAndReboot(Map conf=[:]){
try{
buildHipClangJob(conf)
......@@ -169,14 +260,14 @@ def buildHipClangJobAndReboot(Map conf=[:]){
}
}
def runCKProfiler(Map conf=[:]){
show_node_info()
env.HSA_ENABLE_SDMA=0
checkout scm
def image = "composable_kernels"
def image = "composable_kernels_${params.COMPILER_VERSION}"
def prefixpath = conf.get("prefixpath", "/opt/rocm")
def gpu_arch = conf.get("gpu_arch", "gfx908")
......@@ -184,22 +275,29 @@ def runCKProfiler(Map conf=[:]){
// def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1"
dockerOpts = dockerOpts + " --env HSA_XNACK=1 --env GPU_ARCH='${gpu_arch}' "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' "
if (params.COMPILER_VERSION != "release"){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg GPU_ARCH='${gpu_arch}' "
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
if (params.USE_DOCKERFILE){
try {
retimage = docker.build("${image}", dockerArgs + '.')
//retimage = docker.build("${image}", dockerArgs + '.')
(retimage, image) = getDockerImage(conf)
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES')
{
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
throw new Exception ("GPU not found")
}
else{
echo "GPU is OK"
}
}
}
}
......@@ -208,74 +306,61 @@ def runCKProfiler(Map conf=[:]){
throw e
}
catch(Exception ex) {
retimage = docker.build("${image}", dockerArgs + "--no-cache .")
retimage = docker.build("${image}", dockerArgs + " --no-cache .")
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES')
{
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
throw new Exception ("GPU not found")
}
else{
echo "GPU is OK"
}
}
}
else{
timeout(time: 3, unit: 'HOURS'){
retimage = docker.image('compute-artifactory.amd.com:5000/rocm-plus-docker/framework/compute-rocm-dkms-no-npi-hipclang:9110_ubuntu18.04_py3.6_pytorch_rocm5.0_internal_testing_7ff5b54').pull()
image="b56f8ac0d6ea"
sh "docker images"
}
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 5, unit: 'HOURS')
timeout(time: 24, unit: 'HOURS')
{
cmake_build(conf)
dir("script"){
//run gemm performance tests
def gemm_log = "perf_gemm_${gpu_arch}.log"
sh "rm -f ${gemm_log}"
sh "echo Branch name: ${env.BRANCH_NAME} > ${gemm_log}"
sh "echo Node name: ${NODE_NAME} >> ${gemm_log}"
sh "echo GPU_arch name: ${gpu_arch} >> ${gemm_log}"
sh "rocminfo | grep 'Compute Unit:' >> ${gemm_log} "
sh "hipcc --version | grep -e 'HIP version' >> ${gemm_log}"
sh "/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}"
sh "./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}"
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 name: ${gpu_arch} >> ${resnet_log}"
sh "rocminfo | grep 'Compute Unit:' >> ${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 ${resnet_log} "
if (params.RUN_FULL_QA){
def qa_log = "qa_${gpu_arch}.log"
sh "./run_full_performance_tests.sh 1 QA_${params.COMPILER_VERSION} ${gpu_arch} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N256_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N4_${gpu_arch}.log"
archiveArtifacts "perf_batched_gemm_${gpu_arch}.log"
archiveArtifacts "perf_grouped_gemm_${gpu_arch}.log"
archiveArtifacts "perf_conv_fwd_${gpu_arch}.log"
archiveArtifacts "perf_conv_bwd_data_${gpu_arch}.log"
archiveArtifacts "perf_gemm_bilinear_${gpu_arch}.log"
archiveArtifacts "perf_reduction_${gpu_arch}.log"
// stash perf files to master
stash name: "perf_gemm_${gpu_arch}.log"
stash name: "perf_resnet50_N256_${gpu_arch}.log"
stash name: "perf_resnet50_N4_${gpu_arch}.log"
stash name: "perf_batched_gemm_${gpu_arch}.log"
stash name: "perf_grouped_gemm_${gpu_arch}.log"
stash name: "perf_conv_fwd_${gpu_arch}.log"
stash name: "perf_conv_bwd_data_${gpu_arch}.log"
stash name: "perf_gemm_bilinear_${gpu_arch}.log"
stash name: "perf_reduction_${gpu_arch}.log"
//we will process results on the master node
}
else{
sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${gpu_arch} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N256_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N4_${gpu_arch}.log"
// stash perf files to master
stash name: "perf_gemm_${gpu_arch}.log"
stash name: "perf_resnet50_N256_${gpu_arch}.log"
stash name: "perf_resnet50_N4_${gpu_arch}.log"
//we will process the results on the master node
}
}
}
}
......@@ -283,7 +368,6 @@ def runCKProfiler(Map conf=[:]){
return retimage
}
def runPerfTest(Map conf=[:]){
try{
runCKProfiler(conf)
......@@ -300,16 +384,97 @@ def runPerfTest(Map conf=[:]){
}
}
def process_results(Map conf=[:]){
env.HSA_ENABLE_SDMA=0
checkout scm
def image = "composable_kernels_${params.COMPILER_VERSION}"
def prefixpath = "/opt/rocm"
def gpu_arch = conf.get("gpu_arch", "gfx908")
// Jenkins is complaining about the render group
def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 --env GPU_ARCH='${gpu_arch}' "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='release' "
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
try {
//retimage = docker.build("${image}", dockerArgs + '.')
(retimage, image) = getDockerImage(conf)
}
catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
echo "The job was cancelled or aborted"
throw e
}
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 1, unit: 'HOURS'){
try{
dir("script"){
if (params.RUN_FULL_QA){
// unstash perf files to master
unstash "perf_gemm_${gpu_arch}.log"
unstash "perf_resnet50_N256_${gpu_arch}.log"
unstash "perf_resnet50_N4_${gpu_arch}.log"
unstash "perf_batched_gemm_${gpu_arch}.log"
unstash "perf_grouped_gemm_${gpu_arch}.log"
unstash "perf_conv_fwd_${gpu_arch}.log"
unstash "perf_conv_bwd_data_${gpu_arch}.log"
unstash "perf_gemm_bilinear_${gpu_arch}.log"
unstash "perf_reduction_${gpu_arch}.log"
sh "./process_qa_data.sh ${gpu_arch}"
}
else{
// unstash perf files to master
unstash "perf_gemm_${gpu_arch}.log"
unstash "perf_resnet50_N256_${gpu_arch}.log"
unstash "perf_resnet50_N4_${gpu_arch}.log"
sh "./process_perf_data.sh ${gpu_arch}"
}
}
}
catch(e){
echo "throwing error exception while processing performance test results"
echo 'Exception occurred: ' + e.toString()
throw e
}
}
}
}
//launch develop branch daily at 23:00 in FULL_QA mode
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true''' : ""
pipeline {
agent none
triggers {
parameterizedCron(CRON_SETTINGS)
}
options {
parallelsAlwaysFailFast()
}
parameters {
booleanParam(
name: "USE_DOCKERFILE",
name: "BUILD_DOCKER",
defaultValue: true,
description: "")
description: "Force building docker image (default: true)")
string(
name: 'COMPILER_VERSION',
defaultValue: 'ck-9110',
description: 'Specify which version of compiler to use: ck-9110 (default), release, or amd-stg-open.')
booleanParam(
name: "RUN_FULL_QA",
defaultValue: false,
description: "Select whether to run small set of performance tests (default) or full QA")
booleanParam(
name: "TEST_NODE_PERFORMANCE",
defaultValue: false,
description: "Test the node GPU performance (default: false)")
}
environment{
dbuser = "${dbuser}"
......@@ -319,9 +484,28 @@ pipeline {
dbsshuser = "${dbsshuser}"
dbsshpassword = "${dbsshpassword}"
status_wrapper_creds = "${status_wrapper_creds}"
gerrit_cred="${gerrit_cred}"
DOCKER_BUILDKIT = "1"
}
stages{
stage("Build Docker"){
when {
expression { params.BUILD_DOCKER.toBoolean() }
}
parallel{
stage('Docker /opt/rocm'){
agent{ label rocmnode("nogpu") }
steps{
buildDocker('/opt/rocm')
}
}
}
}
stage("Static checks") {
when {
beforeAgent true
expression { !params.TEST_NODE_PERFORMANCE.toBoolean() }
}
parallel{
// enable after we move from hipcc to hip-clang
// stage('Tidy') {
......@@ -355,6 +539,10 @@ pipeline {
}
stage("Tests")
{
when {
beforeAgent true
expression { !params.TEST_NODE_PERFORMANCE.toBoolean() }
}
parallel
{
stage("Run Tests: gfx908")
......@@ -369,6 +557,10 @@ pipeline {
}
stage("Run Tests: gfx90a")
{
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("gfx90a")}
environment{
setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx90a -O3 " -DBUILD_DEV=On """
......@@ -381,6 +573,10 @@ pipeline {
}
stage("Client App")
{
when {
beforeAgent true
expression { !params.TEST_NODE_PERFORMANCE.toBoolean() }
}
parallel
{
stage("Run Client App")
......@@ -402,6 +598,10 @@ pipeline {
{
stage("Run ckProfiler: gfx908")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && !params.TEST_NODE_PERFORMANCE.toBoolean() }
}
agent{ label rocmnode("gfx908")}
environment{
setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " -DBUILD_DEV=On """
......@@ -412,6 +612,10 @@ pipeline {
}
stage("Run ckProfiler: gfx90a")
{
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() || params.TEST_NODE_PERFORMANCE.toBoolean() }
}
agent{ label rocmnode("gfx90a")}
environment{
setup_args = """ -D CMAKE_CXX_FLAGS="--offload-arch=gfx90a -O3 " -DBUILD_DEV=On """
......@@ -422,6 +626,33 @@ pipeline {
}
}
}
stage("Process Performance Test Results")
{
parallel
{
stage("Process results for gfx908"){
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && !params.TEST_NODE_PERFORMANCE.toBoolean() }
}
agent { label 'mici' }
steps{
process_results(gpu_arch: "gfx908")
}
}
stage("Process results for gfx90a"){
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() || params.TEST_NODE_PERFORMANCE.toBoolean() }
}
agent { label 'mici' }
steps{
process_results(gpu_arch: "gfx90a")
}
}
}
}
/* enable after the cmake file supports packaging
stage("Packages") {
when {
......
README.md
View file @ aa5859e4
......@@ -10,7 +10,7 @@ rocm/tensorflow:rocm5.1-tf2.6-dev \
/bin/bash
```
# Install the new rocm-cmake version
# Install newer version of rocm-cmake
https://github.com/RadeonOpenCompute/rocm-cmake
## Build
......@@ -26,6 +26,7 @@ cmake \
-D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3" \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_INSTALL_PREFIX=${PATH_TO_CK_INSTALL_DIRECTORY} \
..
```
......@@ -47,6 +48,13 @@ Instructions for running each individual examples are under ```example/```
```
Instructions for running ckProfiler are under ```profiler/```
## Install CK
```bash
make install
```
## Using CK as pre-built kernel library
Instructions for using CK as a pre-built kernel library are under ```client_example/```
## Caveat
### Kernel Timing and Verification
......
client_example/01_gemm/CMakeLists.txt 0 → 100644
View file @ aa5859e4
add_executable(client_gemm gemm.cpp)
target_link_libraries(client_gemm PRIVATE composable_kernel::device_operations)
client_example/01_gemm/gemm.cpp 0 → 100644
View file @ aa5859e4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, 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_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm.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 AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
using ADataType = F16;
using BDataType = F16;
using CDataType = F16;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
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(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 7)
{
M = std::stoi(argv[1]);
N = std::stoi(argv[2]);
K = std::stoi(argv[3]);
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideC = std::stoi(argv[6]);
}
else
{
printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
using DeviceOp =
ck::tensor_operation::device::DeviceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
// 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 c_element_op = CElementOp{};
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;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
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: " << 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;
// 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(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
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;
}
return 0;
}
client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
View file @ aa5859e4
......@@ -10,7 +10,7 @@
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/device_gemm_add_add_fastgelu_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
using F16 = ck::half_t;
using F32 = float;
......@@ -27,7 +27,6 @@ using CDEElementOp = AddAddFastGelu;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using D0DataType = F16;
using D1DataType = F16;
using EDataType = F16;
......@@ -111,19 +110,22 @@ int main(int argc, char* argv[])
f_matrix_space_size(M, N, StrideD1, D1Layout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
// add device op instances
const auto op_ptrs = ck::tensor_operation::device::device_gemm_instance::
get_device_gemm_add_add_fastgelu_instances<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>();
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
......@@ -231,6 +233,8 @@ int main(int argc, char* argv[])
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
......
client_example/03_gemm_layernorm/CMakeLists.txt
View file @ aa5859e4
add_executable(gemm_add_add_reduce_normalize gemm_add_add_layernorm.cpp)
target_link_libraries(gemm_add_add_reduce_normalize PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_add_add_reduce_normalize gemm_add_add_layernorm.cpp)
target_link_libraries(client_gemm_add_add_reduce_normalize PRIVATE composable_kernel::device_operations)
client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp
View file @ aa5859e4
......@@ -160,8 +160,9 @@ int main()
ck::index_t StrideC = 1024;
ck::index_t StrideD0 = 1024;
const auto gemm_reduce_ptrs = ck::tensor_operation::device::device_gemm_instance::
get_device_gemm_add_add_mean_squaremean_instances<ADataType,
const auto gemm_reduce_ptrs =
ck::tensor_operation::device::instance::get_device_gemm_add_add_mean_squaremean_instances<
ADataType,
BDataType,
CDataType,
ALayout,
......@@ -169,7 +170,7 @@ int main()
CLayout>();
const auto normalize_ptrs =
ck::tensor_operation::device::get_device_normalize_from_mean_meansquare_instances<
ck::tensor_operation::device::instance::get_device_normalize_from_mean_meansquare_instances<
CDataType,
ReduceDataType,
ReduceDataType,
......
client_example/04_contraction/CMakeLists.txt 0 → 100644
View file @ aa5859e4
add_executable(client_contraction_scale contraction_scale.cpp)
target_link_libraries(client_contraction_scale PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_bilinear contraction_bilinear.cpp)
target_link_libraries(client_contraction_bilinear PRIVATE composable_kernel::device_operations)
client_example/04_contraction/contraction_bilinear.cpp 0 → 100644
View file @ aa5859e4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <numeric>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Bilinear;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F32;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
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(int argc, char* argv[])
{
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// D[M0, M1, N0, N1]
std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float alpha = 1.f;
float beta = 1.f;
if(argc == 1)
{
// use default case
}
else if(argc == 25)
{
const ck::index_t M0 = std::stoi(argv[1]);
const ck::index_t M1 = std::stoi(argv[2]);
const ck::index_t N0 = std::stoi(argv[3]);
const ck::index_t N1 = std::stoi(argv[4]);
const ck::index_t K0 = std::stoi(argv[5]);
const ck::index_t K1 = std::stoi(argv[6]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
d_ms_ns_lengths = {M0, M1, N0, N1};
d_ms_ns_strides = {
std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
alpha = std::stof(argv[23]);
beta = std::stof(argv[24]);
}
else
{
printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg23 to 24: alpha, beta\n");
exit(0);
}
auto f_tensor_space_size = [](auto lengths, auto strides) {
std::size_t space_size = 1;
for(std::size_t i = 0; i < lengths.size(); ++i)
{
space_size += (lengths[i] - 1) * strides[i];
}
return space_size;
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) *
f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
SimpleDeviceMem b_device_buf(sizeof(BDataType) *
f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
SimpleDeviceMem d_device_buf(sizeof(DDataType) *
f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
SimpleDeviceMem e_device_buf(sizeof(EDataType) *
f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>;
// 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{alpha, beta};
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;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr =
op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
e_ms_ns_lengths.begin() + NumDimM,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
e_ms_ns_lengths.begin() + NumDimM + NumDimN,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
a_ms_ks_lengths.begin() + NumDimM + NumDimK,
ck::index_t{1},
std::multiplies<ck::index_t>{});
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(DDataType) * M * N + sizeof(EDataType) * 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: " << 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/04_contraction/contraction_scale.cpp 0 → 100644
View file @ aa5859e4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <numeric>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Scale;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
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(int argc, char* argv[])
{
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float scale = 1.f;
if(argc == 1)
{
// use default case
}
else if(argc == 20)
{
const ck::index_t M0 = std::stoi(argv[1]);
const ck::index_t M1 = std::stoi(argv[2]);
const ck::index_t N0 = std::stoi(argv[3]);
const ck::index_t N1 = std::stoi(argv[4]);
const ck::index_t K0 = std::stoi(argv[5]);
const ck::index_t K1 = std::stoi(argv[6]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
scale = std::stof(argv[19]);
}
else
{
printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg19: scale\n");
exit(0);
}
auto f_tensor_space_size = [](auto lengths, auto strides) {
std::size_t space_size = 1;
for(std::size_t i = 0; i < lengths.size(); ++i)
{
space_size += (lengths[i] - 1) * strides[i];
}
return space_size;
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) *
f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
SimpleDeviceMem b_device_buf(sizeof(BDataType) *
f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
SimpleDeviceMem e_device_buf(sizeof(EDataType) *
f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>;
// 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{scale};
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;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{},
e_device_buf.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 0>{},
std::array<std::vector<ck::index_t>, 0>{},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
e_ms_ns_lengths.begin() + NumDimM,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
e_ms_ns_lengths.begin() + NumDimM + NumDimN,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
a_ms_ks_lengths.begin() + NumDimM + NumDimK,
ck::index_t{1},
std::multiplies<ck::index_t>{});
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(EDataType) * 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: " << 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 @ aa5859e4
......@@ -6,5 +6,7 @@ find_package(composable_kernel 1.0.0 COMPONENTS device_operations)
find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}")
add_subdirectory(01_gemm)
add_subdirectory(02_gemm_add_add_fastgelu)
add_subdirectory(03_gemm_layernorm)
add_subdirectory(04_contraction)
client_example/README.md
View file @ aa5859e4
##
Client application links to CK library, and therefore CK library needs to be installed before building client applications.
## Docker script
```bash
docker run \
-it \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm5.1-tf2.6-dev \
/bin/bash
```
## Build
```bash
......@@ -22,7 +11,7 @@ cd client_example/build
```bash
cmake \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_PREFIX_PATH="/opt/rocm;${PATH_TO_CK_INSTALL_DIRECTORY}" \
..
```
......
cmake/googletest.cmake
View file @ aa5859e4
......@@ -20,6 +20,7 @@ list(APPEND GTEST_CMAKE_CXX_FLAGS
-Wno-unused-member-function
-Wno-comma
-Wno-old-style-cast
-Wno-deprecated
)
message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
......
example/01_gemm/CMakeLists.txt
View file @ aa5859e4
......@@ -4,5 +4,6 @@ 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_bf16 gemm_xdl_bf16.cpp)
add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.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
View file @ aa5859e4
......@@ -12,9 +12,9 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is>
......@@ -142,9 +142,9 @@ int main(int argc, char* argv[])
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());
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
......
example/01_gemm/gemm_dl_fp32.cpp
View file @ aa5859e4
......@@ -12,9 +12,9 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is>
......@@ -141,9 +141,9 @@ int main(int argc, char* argv[])
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());
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
......
example/01_gemm/gemm_dl_int8.cpp
View file @ aa5859e4
......@@ -12,9 +12,9 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is>
......@@ -139,9 +139,9 @@ int main(int argc, char* argv[])
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());
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
......
example/01_gemm/gemm_xdl_bf16.cpp
View file @ aa5859e4
......@@ -11,9 +11,9 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -170,9 +170,9 @@ int main(int argc, char* argv[])
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
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());
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
......
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