Skip to content

GitLab

Commit 9c0811f3 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents ded0d83d 3528a523
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2085 additions and 310 deletions
Dockerfile
View file @ 9c0811f3
...@@ -23,6 +23,11 @@ RUN if [ "$ROCMVERSION" != "6.3" ]; then \ ...@@ -23,6 +23,11 @@ RUN if [ "$ROCMVERSION" != "6.3" ]; then \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \ wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \ sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \ sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "6.3" ] && [ "$compiler_version" = "rc1" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.3.0.1-20.04-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.3.0.1-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.3.0.1 rel-5 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2033700; \
fi fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
...@@ -130,6 +135,8 @@ ENV compiler_commit=$compiler_commit ...@@ -130,6 +135,8 @@ ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'" RUN sh -c "echo compiler version = '$compiler_version'"
RUN sh -c "echo compiler commit = '$compiler_commit'" RUN sh -c "echo compiler commit = '$compiler_commit'"
ARG DISABLE_CACHE=0
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline-open" ] ) && [ "$compiler_commit" = "" ]; then \ RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline-open" ] ) && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \ git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
cd llvm-project && mkdir build && cd build && \ cd llvm-project && mkdir build && cd build && \
......
Jenkinsfile
View file @ 9c0811f3
...@@ -94,13 +94,21 @@ def getDockerImage(Map conf=[:]){ ...@@ -94,13 +94,21 @@ def getDockerImage(Map conf=[:]){
env.DOCKER_BUILDKIT=1 env.DOCKER_BUILDKIT=1
def prefixpath = conf.get("prefixpath", "/opt/rocm") def prefixpath = conf.get("prefixpath", "/opt/rocm")
def no_cache = conf.get("no_cache", false) def no_cache = conf.get("no_cache", false)
def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${prefixpath} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' " def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${prefixpath} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' --build-arg DISABLE_CACHE='git rev-parse ${params.COMPILER_VERSION}' "
if(no_cache) if(no_cache)
{ {
dockerArgs = dockerArgs + " --no-cache " dockerArgs = dockerArgs + " --no-cache "
} }
echo "Docker Args: ${dockerArgs}" echo "Docker Args: ${dockerArgs}"
def image = getDockerImageName() def image
if ( params.BUILD_LEGACY_OS && conf.get("docker_name", "") != "" ){
image = conf.get("docker_name", "")
echo "Using legacy docker: ${image}"
}
else{
image = getDockerImageName()
echo "Using default docker: ${image}"
}
//Check if image exists //Check if image exists
def retimage def retimage
try try
...@@ -124,8 +132,10 @@ def buildDocker(install_prefix){ ...@@ -124,8 +132,10 @@ def buildDocker(install_prefix){
checkout scm checkout scm
def image_name = getDockerImageName() def image_name = getDockerImageName()
echo "Building Docker for ${image_name}" echo "Building Docker for ${image_name}"
def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' " def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' --build-arg DISABLE_CACHE='git rev-parse ${params.COMPILER_VERSION}' "
if(params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
dockerArgs = dockerArgs + " --no-cache "
}
echo "Build Args: ${dockerArgs}" echo "Build Args: ${dockerArgs}"
try{ try{
if(params.BUILD_DOCKER){ if(params.BUILD_DOCKER){
...@@ -259,6 +269,7 @@ def cmake_build(Map conf=[:]){ ...@@ -259,6 +269,7 @@ def cmake_build(Map conf=[:]){
""") """)
sh cmd3 sh cmd3
} }
// reduce parallelism when compiling, clang uses too much memory // reduce parallelism when compiling, clang uses too much memory
def nt = nthreads() def nt = nthreads()
def cmd def cmd
...@@ -273,7 +284,7 @@ def cmake_build(Map conf=[:]){ ...@@ -273,7 +284,7 @@ def cmake_build(Map conf=[:]){
} }
else{ else{
setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ") setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j${nt} ${config_targets}") build_cmd = conf.get("build_cmd", "${build_envs} make -j${nt} ${config_targets}")
} }
cmd = conf.get("cmd", """ cmd = conf.get("cmd", """
${setup_cmd} ${setup_cmd}
...@@ -292,8 +303,8 @@ def cmake_build(Map conf=[:]){ ...@@ -292,8 +303,8 @@ def cmake_build(Map conf=[:]){
dir("build"){ dir("build"){
//build CK //build CK
sh cmd sh cmd
//run tests //run tests except when NO_CK_BUILD or BUILD_LEGACY_OS are set
if(!setup_args.contains("NO_CK_BUILD")){ if(!setup_args.contains("NO_CK_BUILD") && !params.BUILD_LEGACY_OS){
if (setup_args.contains("gfx90a") && params.NINJA_BUILD_TRACE){ if (setup_args.contains("gfx90a") && params.NINJA_BUILD_TRACE){
sh "/ninjatracing/ninjatracing .ninja_log > ck_build_trace.json" sh "/ninjatracing/ninjatracing .ninja_log > ck_build_trace.json"
archiveArtifacts "ck_build_trace.json" archiveArtifacts "ck_build_trace.json"
...@@ -330,7 +341,15 @@ def buildHipClangJob(Map conf=[:]){ ...@@ -330,7 +341,15 @@ def buildHipClangJob(Map conf=[:]){
env.HSA_ENABLE_SDMA=0 env.HSA_ENABLE_SDMA=0
checkout scm checkout scm
def image = getDockerImageName() def image
if ( params.BUILD_LEGACY_OS && conf.get("docker_name", "") != "" ){
image = conf.get("docker_name", "")
echo "Using legacy docker: ${image}"
}
else{
image = getDockerImageName()
echo "Using default docker: ${image}"
}
def prefixpath = conf.get("prefixpath", "/opt/rocm") def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group // Jenkins is complaining about the render group
...@@ -512,7 +531,16 @@ def Build_CK(Map conf=[:]){ ...@@ -512,7 +531,16 @@ def Build_CK(Map conf=[:]){
env.DOCKER_BUILDKIT=1 env.DOCKER_BUILDKIT=1
checkout scm checkout scm
def image = getDockerImageName() def image
if ( params.BUILD_LEGACY_OS && conf.get("docker_name", "") != "" ){
image = conf.get("docker_name", "")
echo "Using legacy docker: ${image}"
}
else{
image = getDockerImageName()
echo "Using default docker: ${image}"
}
def prefixpath = conf.get("prefixpath", "/opt/rocm") def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group // Jenkins is complaining about the render group
...@@ -524,6 +552,9 @@ def Build_CK(Map conf=[:]){ ...@@ -524,6 +552,9 @@ def Build_CK(Map conf=[:]){
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){ if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' " dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
} }
if(params.BUILD_LEGACY_OS){
dockerOpts = dockerOpts + " --env LD_LIBRARY_PATH='/opt/Python-3.8.13/lib' "
}
def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3') def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3')
def render_id = sh(returnStdout: true, script: 'getent group render | cut -d: -f3') def render_id = sh(returnStdout: true, script: 'getent group render | cut -d: -f3')
dockerOpts = dockerOpts + " --group-add=${video_id} --group-add=${render_id} " dockerOpts = dockerOpts + " --group-add=${video_id} --group-add=${render_id} "
...@@ -703,11 +734,12 @@ def process_results(Map conf=[:]){ ...@@ -703,11 +734,12 @@ def process_results(Map conf=[:]){
} }
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2; RUN_CK_TILE_TESTS=true CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true 0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true 0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true 0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : "" 0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
0 13 * * * % BUILD_LEGACY_OS=true ''' : ""
pipeline { pipeline {
agent none agent none
...@@ -775,9 +807,13 @@ pipeline { ...@@ -775,9 +807,13 @@ pipeline {
defaultValue: false, defaultValue: false,
description: "Run the grouped conv large cases tests (default: OFF)") description: "Run the grouped conv large cases tests (default: OFF)")
booleanParam( booleanParam(
name: "RUN_CK_TILE_TESTS", name: "RUN_CK_TILE_FMHA_TESTS",
defaultValue: false, defaultValue: false,
description: "Run the ck_tile tests (default: OFF)") description: "Run the ck_tile FMHA tests (default: OFF)")
booleanParam(
name: "RUN_CK_TILE_GEMM_TESTS",
defaultValue: false,
description: "Run the ck_tile GEMM tests (default: OFF)")
booleanParam( booleanParam(
name: "BUILD_INSTANCES_ONLY", name: "BUILD_INSTANCES_ONLY",
defaultValue: false, defaultValue: false,
...@@ -790,6 +826,10 @@ pipeline { ...@@ -790,6 +826,10 @@ pipeline {
name: "NINJA_BUILD_TRACE", name: "NINJA_BUILD_TRACE",
defaultValue: false, defaultValue: false,
description: "Generate a ninja build trace (default: OFF)") description: "Generate a ninja build trace (default: OFF)")
booleanParam(
name: "BUILD_LEGACY_OS",
defaultValue: false,
description: "Try building CK with legacy OS dockers: RHEL8 and SLES15 (default: OFF)")
} }
environment{ environment{
dbuser = "${dbuser}" dbuser = "${dbuser}"
...@@ -894,15 +934,15 @@ pipeline { ...@@ -894,15 +934,15 @@ pipeline {
} }
} }
} }
stage("Run CK_TILE Tests") stage("Run CK_TILE_FMHA Tests")
{ {
parallel parallel
{ {
stage("Run CK_TILE Tests on gfx90a") stage("Run CK_TILE_FMHA Tests on gfx90a")
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.RUN_CK_TILE_TESTS.toBoolean() } expression { params.RUN_CK_TILE_FMHA_TESTS.toBoolean() }
} }
agent{ label rocmnode("gfx90a") } agent{ label rocmnode("gfx90a") }
environment{ environment{
...@@ -917,11 +957,11 @@ pipeline { ...@@ -917,11 +957,11 @@ pipeline {
cleanWs() cleanWs()
} }
} }
stage("Run CK_TILE Tests on gfx942") stage("Run CK_TILE_FMHA Tests on gfx942")
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.RUN_CK_TILE_TESTS.toBoolean() } expression { params.RUN_CK_TILE_FMHA_TESTS.toBoolean() }
} }
agent{ label rocmnode("gfx942") } agent{ label rocmnode("gfx942") }
environment{ environment{
...@@ -938,21 +978,103 @@ pipeline { ...@@ -938,21 +978,103 @@ pipeline {
} }
} }
} }
stage("Run CK_TILE_GEMM Tests")
{
parallel
{
stage("Run CK_TILE_GEMM Tests on gfx90a")
{
when {
beforeAgent true
expression { params.RUN_CK_TILE_GEMM_TESTS.toBoolean() }
}
agent{ label rocmnode("gfx90a") }
environment{
setup_args = "NO_CK_BUILD"
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx90a && \
make -j64 tile_example_gemm_basic && \
cd ../ &&
example/ck_tile/03_gemm/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx90a """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
cleanWs()
}
}
stage("Run CK_TILE_GEMM Tests on gfx942")
{
when {
beforeAgent true
expression { params.RUN_CK_TILE_GEMM_TESTS.toBoolean() }
}
agent{ label rocmnode("gfx942") }
environment{
setup_args = "NO_CK_BUILD"
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx942 && \
make -j64 tile_example_gemm_basic && \
cd ../ &&
example/ck_tile/03_gemm/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx942 """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
cleanWs()
}
}
}
}
stage("Build CK and run Tests") stage("Build CK and run Tests")
{ {
parallel parallel
{ {
stage("Build CK with RHEL8")
{
when {
beforeAgent true
expression { params.BUILD_LEGACY_OS.toBoolean() }
}
agent{ label rocmnode("gfx90a") }
environment{
def docker_name = "${env.CK_DOCKERHUB_PRIVATE}:ck_rhel8_rocm6.3"
setup_args = """ -DGPU_TARGETS="gfx942" \
-DCMAKE_CXX_FLAGS=" -O3 " \
-DCK_USE_ALTERNATIVE_PYTHON=/opt/Python-3.8.13/bin/python3.8 """
execute_args = " "
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: " ", no_reboot:true, build_type: 'Release', docker_name: docker_name)
cleanWs()
}
}
stage("Build CK with SLES15")
{
when {
beforeAgent true
expression { params.BUILD_LEGACY_OS.toBoolean() }
}
agent{ label rocmnode("gfx90a") }
environment{
def docker_name = "${env.CK_DOCKERHUB_PRIVATE}:ck_sles15_rocm6.3"
setup_args = """ -DGPU_TARGETS="gfx942" \
-DCMAKE_CXX_FLAGS=" -O3 " \
-DCK_USE_ALTERNATIVE_PYTHON=/opt/Python-3.8.13/bin/python3.8 """
execute_args = " "
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: " ", no_reboot:true, build_type: 'Release', docker_name: docker_name)
cleanWs()
}
}
stage("Build CK for all gfx9 targets") stage("Build CK for all gfx9 targets")
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() } expression { params.RUN_FULL_QA.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx90a") } agent{ label rocmnode("gfx90a") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install \ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install \
-DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" \ -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" \
-DCMAKE_EXE_LINKER_FLAGS=" -L ${env.WORKSPACE}/script -T hip_fatbin_insert " \
-DCMAKE_CXX_FLAGS=" -O3 " """ -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \ cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
...@@ -969,7 +1091,7 @@ pipeline { ...@@ -969,7 +1091,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() } expression { params.RUN_FULL_QA.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx942") } agent{ label rocmnode("gfx942") }
environment{ environment{
...@@ -989,7 +1111,7 @@ pipeline { ...@@ -989,7 +1111,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() } expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx90a") } agent{ label rocmnode("gfx90a") }
environment{ environment{
...@@ -1009,7 +1131,7 @@ pipeline { ...@@ -1009,7 +1131,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.BUILD_INSTANCES_ONLY.toBoolean() && !params.RUN_FULL_QA.toBoolean() } expression { params.BUILD_INSTANCES_ONLY.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx90a") } agent{ label rocmnode("gfx90a") }
environment{ environment{
...@@ -1028,7 +1150,7 @@ pipeline { ...@@ -1028,7 +1150,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() } expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx1030") } agent{ label rocmnode("gfx1030") }
environment{ environment{
...@@ -1048,7 +1170,7 @@ pipeline { ...@@ -1048,7 +1170,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() } expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx1101") } agent{ label rocmnode("gfx1101") }
environment{ environment{
...@@ -1068,7 +1190,7 @@ pipeline { ...@@ -1068,7 +1190,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() } expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent{ label rocmnode("gfx1201") } agent{ label rocmnode("gfx1201") }
environment{ environment{
...@@ -1095,7 +1217,7 @@ pipeline { ...@@ -1095,7 +1217,7 @@ pipeline {
{ {
when { when {
beforeAgent true beforeAgent true
expression { params.RUN_PERFORMANCE_TESTS.toBoolean() } expression { params.RUN_PERFORMANCE_TESTS.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
options { retry(1) } options { retry(1) }
agent{ label rocmnode("gfx90a")} agent{ label rocmnode("gfx90a")}
...@@ -1116,7 +1238,7 @@ pipeline { ...@@ -1116,7 +1238,7 @@ pipeline {
stage("Process results"){ stage("Process results"){
when { when {
beforeAgent true beforeAgent true
expression { params.RUN_PERFORMANCE_TESTS.toBoolean() } expression { params.RUN_PERFORMANCE_TESTS.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
} }
agent { label 'mici' } agent { label 'mici' }
steps{ steps{
......
docs/sphinx/requirements.in
View file @ 9c0811f3
rocm-docs-core==1.7.2 rocm-docs-core==1.8.1
sphinxcontrib-bibtex==2.6.2 sphinxcontrib-bibtex==2.6.3
docs/sphinx/requirements.txt
View file @ 9c0811f3
...@@ -103,7 +103,7 @@ requests==2.32.3 ...@@ -103,7 +103,7 @@ requests==2.32.3
# via # via
# pygithub # pygithub
# sphinx # sphinx
rocm-docs-core==1.7.2 rocm-docs-core==1.8.1
# via -r requirements.in # via -r requirements.in
six==1.16.0 six==1.16.0
# via pybtex # via pybtex
...@@ -137,7 +137,7 @@ sphinx-notfound-page==1.0.3 ...@@ -137,7 +137,7 @@ sphinx-notfound-page==1.0.3
# via rocm-docs-core # via rocm-docs-core
sphinxcontrib-applehelp==2.0.0 sphinxcontrib-applehelp==2.0.0
# via sphinx # via sphinx
sphinxcontrib-bibtex==2.6.2 sphinxcontrib-bibtex==2.6.3
# via -r requirements.in # via -r requirements.in
sphinxcontrib-devhelp==2.0.0 sphinxcontrib-devhelp==2.0.0
# via sphinx # via sphinx
......
example/01_gemm/run_gemm_example.inc
View file @ 9c0811f3
...@@ -305,6 +305,14 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -305,6 +305,14 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
} }
#endif #endif
} }
else
{
// When the Problem Type and Problem Size does not fit.
std::cerr << gemm.GetTypeString() << ": the instance does not support the problem config."
<< std::endl;
return true;
}
std::size_t flop = 2_uz * M * N * K; std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype = std::size_t num_btype =
......
example/01_gemm/run_gemm_example_v2.inc
View file @ 9c0811f3
...@@ -161,18 +161,6 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -161,18 +161,6 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
} }
#if 0
printf("B matrix:\n");
for (int in = 0; in < N; in++)
{
for (int ik = 0; ik < K; ik++)
{
printf("%02x ", *(reinterpret_cast<uint8_t*>(&b_k_n(ik,in))));
if(ik%8==7) printf("|");
}
printf("\n");
}
#endif
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); 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{})); Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
...@@ -272,7 +260,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -272,7 +260,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
if(config.time_kernel) if(config.time_kernel)
{ {
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel}); ave_time =
invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 0, 5, 10, true, 4});
std::size_t flop = 2_uz * M * N * K; std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype = std::size_t num_btype =
......
example/ck_tile/03_gemm/CMakeLists.txt 0 → 100644
View file @ 9c0811f3
set(CMAKE_BUILD_TYPE Debug)
add_executable(tile_example_gemm_basic EXCLUDE_FROM_ALL gemm_basic.cpp)
\ No newline at end of file
example/ck_tile/03_gemm/README.md 0 → 100644
View file @ 9c0811f3
# GEMM Matrix Multiplication
This folder contains example for GEMM using ck_tile tile-programming implementation. Currently, it only supports the basic feature of the CK Tile GEMM, but creates the placeholders for the future support on different GEMM pipeline and different GEMM modules. In the near future, we will gradually migrate all the GEMM features from old CK to CK Tile.
## build
```
# in the root of ck_tile
mkdir build && cd build
sh ../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_gemm_basic -j
```
This will result in an executable `build/bin/tile_example_gemm_basic`
## example
```
args:
-m m dimension (default:3328)
-n m dimension (default:4096)
-k k dimension (default:64)
-e epsilon (default:1e-5)
-v cpu validation or not (default:1)
-prec precision (default:fp16)
```
example/ck_tile/03_gemm/gemm_basic.cpp 0 → 100644
View file @ 9c0811f3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_basic.hpp"
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("b", "1", "batch size")
.insert("m", "1024", "m dimension")
.insert("n", "2048", "n dimension")
.insert("k", "64", "k dimension")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("e", "1e-5", "Absolute error tolerance")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "10", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename LayoutA,
typename LayoutB,
typename LayoutC,
typename PipelineProblem,
typename GemmPipeline,
typename GemmShape>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
{
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr int kBlockPerCu = 1;
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadA, kPadB>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel =
ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue, LayoutA, LayoutB, LayoutC>;
auto kargs = Kernel::MakeKargs(args.p_a,
args.p_b,
args.p_c,
args.epsilon,
args.M,
args.N,
args.K,
args.stride_A,
args.stride_B,
args.stride_C);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize();
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
template <typename DataType,
typename LayoutA,
typename LayoutB,
typename LayoutC,
typename PipelineProblem,
typename GemmPipeline,
typename GemmShape>
float invoke_gemm(ck_tile::DeviceMem& a_buf,
ck_tile::DeviceMem& b_buf,
ck_tile::DeviceMem& c_buf,
const ck_tile::ArgParser& arg_parser)
{
std::string data_type = arg_parser.get_str("prec");
if(data_type != DataTypeTraits<DataType>::name)
{
std::cerr << "Data type mismatch: expected " << DataTypeTraits<DataType>::name << ", got "
<< data_type << std::endl;
return -1; // Or handle the error appropriately
}
float epsilon = arg_parser.get_float("e");
ck_tile::index_t batch_size = arg_parser.get_int("b");
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
ck_tile::index_t stride_a = arg_parser.get_int("stride_a");
ck_tile::index_t stride_b = arg_parser.get_int("stride_b");
ck_tile::index_t stride_c = arg_parser.get_int("stride_c");
gemm_basic_args args;
args.p_a = a_buf.GetDeviceBuffer();
args.p_b = b_buf.GetDeviceBuffer();
args.p_c = c_buf.GetDeviceBuffer();
args.epsilon = epsilon;
args.kbatch = batch_size;
args.M = M;
args.N = N;
args.K = K;
// Only set stride_M and stride_N if they are non-zero and not equal to K.
if(stride_a != 0)
{
args.stride_A = stride_a;
}
else
{
args.stride_A = [&]() {
if constexpr(std::is_same_v<LayoutA, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
return M;
}
else
{
return K;
}
}();
}
if(stride_b != 0)
{
args.stride_B = stride_b;
}
else
{
args.stride_B = [&]() {
if constexpr(std::is_same_v<LayoutB, ck_tile::tensor_layout::gemm::RowMajor>)
{
return N;
}
else
{
return K;
}
}();
}
if(stride_c != 0)
{
args.stride_C = stride_c;
}
else
{
args.stride_C = [&]() {
if constexpr(std::is_same_v<LayoutC, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
return M;
}
else
{
return N;
}
}();
}
float ave_time = gemm_calc<LayoutA, LayoutB, LayoutC, PipelineProblem, GemmPipeline, GemmShape>(
args, ck_tile::stream_config{nullptr, true});
std::size_t num_byte =
sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "The overall perfomance of the GEMM with "
<< "[" << data_type << "]"
<< "batch size: " << batch_size << ". m:" << M << ", n:" << N << ", k:" << K
<< " is: \n";
std::cout << "Running time: " << ave_time << "ms, Throughput " << gb_per_sec << "GB/s \n"
<< std::flush;
return ave_time;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
// The Matrix Multiplication goes with Matrix A (M, K), Matrix B (N, K) = Matrix C (M, N).
using matrix_a_layout = ck_tile::tensor_layout::gemm::RowMajor;
using matrix_b_layout = ck_tile::tensor_layout::gemm::ColumnMajor;
using matrix_c_layout = ck_tile::tensor_layout::gemm::RowMajor;
// host verify
std::vector<int> a_dimensions =
(std::is_same_v<matrix_a_layout, ck_tile::tensor_layout::gemm::RowMajor>)
? std::vector<int>{M, K}
: std::vector<int>{K, M};
std::vector<int> b_dimensions =
(std::is_same_v<matrix_b_layout, ck_tile::tensor_layout::gemm::ColumnMajor>)
? std::vector<int>{N, K}
: std::vector<int>{K, N};
std::vector<int> c_dimensions =
(std::is_same_v<matrix_c_layout, ck_tile::tensor_layout::gemm::RowMajor>)
? std::vector<int>{M, N}
: std::vector<int>{N, M};
ck_tile::HostTensor<ADataType> a_host(a_dimensions);
ck_tile::HostTensor<BDataType> b_host(b_dimensions);
ck_tile::HostTensor<CDataType> c_host_ref(c_dimensions);
ck_tile::HostTensor<CDataType> c_host_dev(c_dimensions);
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_host);
ck_tile::DeviceMem a_buf(a_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_buf(b_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_buf(c_host_dev.get_element_space_size_in_bytes());
a_buf.ToDevice(a_host.data());
b_buf.ToDevice(b_host.data());
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
// This part comes from the Codegen
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using CodegenPipelineProblem = ck_tile::BlockGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenGemmShape,
kPadA,
kPadB,
kPadC>;
using CodegenGemmPipeline = ck_tile::BlockGemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
invoke_gemm<ck_tile::half_t,
matrix_a_layout,
matrix_b_layout,
matrix_c_layout,
CodegenPipelineProblem,
CodegenGemmPipeline,
CodegenGemmShape>(a_buf, b_buf, c_buf, arg_parser);
c_buf.FromDevice(c_host_dev.data());
bool pass_cpu = true;
if(arg_parser.get_int("v") == 1)
{
// ToDo: Will Add the Element Op (bias) verification in the future.
ck_tile::reference_gemm<ADataType,
BDataType,
AccDataType,
CDataType,
matrix_a_layout,
matrix_b_layout,
matrix_c_layout>(a_host, b_host, c_host_ref);
pass_cpu = ck_tile::check_err(c_host_dev, c_host_ref);
std::cout << "The CPU veification result is:" << (pass_cpu ? "correct" : "fail")
<< std::flush;
}
bool pass_gpu = true;
if(arg_parser.get_int("v") == 2)
{
ck_tile::index_t stride_a = arg_parser.get_int("stride_a");
ck_tile::index_t stride_b = arg_parser.get_int("stride_b");
ck_tile::index_t stride_c = arg_parser.get_int("stride_c");
if(stride_a == 0)
{
if constexpr(std::is_same_v<matrix_a_layout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
stride_a = M;
}
else
{
stride_a = K;
}
}
if(stride_b == 0)
{
if constexpr(std::is_same_v<matrix_b_layout, ck_tile::tensor_layout::gemm::RowMajor>)
{
stride_b = N;
}
else
{
stride_b = K;
}
}
if(stride_c == 0)
{
if constexpr(std::is_same_v<matrix_c_layout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
stride_c = M;
}
else
{
stride_c = N;
}
}
ck_tile::HostTensor<CDataType> c_host_gpu_ref(c_dimensions);
ck_tile::DeviceMem c_gpu_buf(c_host_gpu_ref.get_element_space_size_in_bytes());
ck_tile::reference_gemm_gpu<ADataType, BDataType, AccDataType, CDataType>(
a_buf, b_buf, c_gpu_buf, M, N, K, stride_a, stride_b, stride_c);
c_buf.FromDevice(c_host_gpu_ref.data());
pass_gpu = ck_tile::check_err(c_host_dev, c_host_gpu_ref);
std::cout << "The GPU veification result is: " << (pass_gpu ? "correct" : "fail")
<< std::flush;
}
std::cout << std::endl << std::flush;
return !pass_gpu;
}
example/ck_tile/03_gemm/gemm_basic.hpp 0 → 100644
View file @ 9c0811f3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/host.hpp"
#include <string>
template <typename DataType>
struct GemmBasicTypeConfig;
template <>
struct GemmBasicTypeConfig<ck_tile::half_t>
{
using ADataType = ck_tile::half_t;
using BDataType = ck_tile::half_t;
using AccDataType = float;
using CDataType = ck_tile::half_t; // type convert
// ToDo: Add more bias config to support different categories of GEMM.
};
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<float>
{
static constexpr const char* name = "fp32";
};
template <>
struct DataTypeTraits<double>
{
static constexpr const char* name = "fp64";
};
template <>
struct DataTypeTraits<ck_tile::half_t>
{
static constexpr const char* name = "fp16";
};
using Types = GemmBasicTypeConfig<ck_tile::half_t>;
// Specific type aliases for easy access
using ADataType = Types::ADataType;
using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType;
struct gemm_basic_args
{
const void* p_a;
const void* p_b;
void* p_c;
float epsilon;
ck_tile::index_t kbatch;
ck_tile::index_t M;
ck_tile::index_t N;
ck_tile::index_t K;
ck_tile::index_t stride_A;
ck_tile::index_t stride_B;
ck_tile::index_t stride_C;
};
// host API
float gemm_calc(gemm_basic_args args, const ck_tile::stream_config& s);
example/ck_tile/03_gemm/script/run_full_test.sh 0 → 100755
View file @ 9c0811f3
#!/bin/bash
#
# in order to run this script you'd first need to build the tile_example_gemm executables in ../build/bin/
#
# run the script as "./run_full_test.sh <tag for your test environment> <branch name> <host name> <gpu_arch>
# input arguments:
# environment tag : a string describing the specifics of your test environment
# branch name : name of the branch in git repo (git status | grep -e 'On branch')
# host name : $hostname
# gpu architecture: e.g., gfx90a, or gfx942, etc.
# get the command line arguments:
export env_type=$1
echo 'Environment type: ' $env_type
export branch=$2
echo 'Branch name: ' $branch
export host_name=$3
echo 'Host name: ' $host_name
export GPU_arch=$4
echo 'GPU_arch: ' $GPU_arch
# run verification tests
example/ck_tile/03_gemm/script/smoke_test.sh
# We do not have a performance benchmark for gemm yet. Will add it in the future.
\ No newline at end of file
example/ck_tile/03_gemm/script/smoke_test.sh 0 → 100755
View file @ 9c0811f3
#!/bin/bash
EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
KNAME=1
export CK_WARMUP=0
export CK_REPEAT=1
COMMON_ARGS='-v=2 -warmup=0 -repeat=1'
run_fp16_tests() {
for batch in 1 2; do
for m in 128 1024; do
for n in 128 2048; do
for k in 32 64; do
$EXE -b=$batch -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -e=1e-5 -prec=fp16 $COMMON_ARGS
if [ $? -eq 0 ]; then
echo "Success: Test with batch=$batch, m=$m, n=$n, k=$k executed successfully."
else
echo "Error: Test with batch=$batch, m=$m, n=$n, k=$k failed to execute properly."
# Optionally, exit or break if you need to halt further execution
# exit 1
fi
done
done
done
done
}
set -x
run_fp16_tests
set +x
\ No newline at end of file
example/ck_tile/CMakeLists.txt
View file @ 9c0811f3
...@@ -4,3 +4,4 @@ include_directories(AFTER ...@@ -4,3 +4,4 @@ include_directories(AFTER
add_subdirectory(01_fmha) add_subdirectory(01_fmha)
add_subdirectory(02_layernorm2d) add_subdirectory(02_layernorm2d)
add_subdirectory(03_gemm)
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp
View file @ 9c0811f3
...@@ -446,7 +446,9 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave, ...@@ -446,7 +446,9 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
}); });
}); });
}); });
__builtin_amdgcn_sched_barrier(0); // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
// latency
// __builtin_amdgcn_sched_barrier(0);
} }
} }
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp
View file @ 9c0811f3
...@@ -171,6 +171,16 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK<ALayo ...@@ -171,6 +171,16 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK<ALayo
Argument arg_ = arg; Argument arg_ = arg;
const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
auto size_a_buffer =
a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
auto size_b_buffer =
b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N( const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N(
arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs); arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs);
...@@ -179,11 +189,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK<ALayo ...@@ -179,11 +189,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK<ALayo
DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType); DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType);
}); });
ck::utility::RotatingMemWrapperMultiD<Argument, DsDataType> rotating_mem( ck::utility::RotatingMemWrapperMultiD<Argument, DsDataType> rotating_mem(
arg_, arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer, DsSize);
stream_config.rotating_count,
arg_.M * arg_.K * sizeof(ADataType),
arg_.K * arg_.N * sizeof(BDataType),
DsSize);
rotating_mem.Print(); rotating_mem.Print();
auto run_flush_cache = [&]() { auto run_flush_cache = [&]() {
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp
View file @ 9c0811f3
...@@ -155,11 +155,19 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout, ...@@ -155,11 +155,19 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
if(stream_config.flush_cache) if(stream_config.flush_cache)
{ {
Argument arg_ = arg; Argument arg_ = arg;
const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
auto size_a_buffer =
a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
auto size_b_buffer =
b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
ck::utility::RotatingMemWrapper<Argument> rotating_mem( ck::utility::RotatingMemWrapper<Argument> rotating_mem(
arg_, arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
stream_config.rotating_count,
arg_.M * arg_.K * sizeof(ADataType),
arg_.K * arg_.N * sizeof(BDataType));
rotating_mem.Print(); rotating_mem.Print();
auto run_flush_cache = [&]() { auto run_flush_cache = [&]() {
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp
View file @ 9c0811f3
...@@ -15,6 +15,7 @@ ...@@ -15,6 +15,7 @@
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
#include "ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm.hpp" #include "ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm.hpp"
#include "ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp" #include "ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp"
#include "ck/tensor_operation/operator_transform/transform_conv_ngchw_to_nhwgc.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp" #include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
...@@ -22,7 +23,6 @@ ...@@ -22,7 +23,6 @@
#include <ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp> #include <ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp>
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
...@@ -257,6 +257,19 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -257,6 +257,19 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
KPerBlock / K1Number, KPerBlock / K1Number,
ConvBackwardWeightSpecialization>{}; ConvBackwardWeightSpecialization>{};
static constexpr index_t ClusterLengthMPerBlock =
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock::At(1);
static constexpr index_t ClusterLengthNPerBlock =
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock::At(3);
static constexpr auto conv_ngchw_to_nhwgc_transformer =
TransformConvNGCHWToNHWGC<InLayout,
WeiLayout,
OutLayout,
NDimSpatial,
MPerBlock / ClusterLengthMPerBlock,
NPerBlock / ClusterLengthNPerBlock>{};
static constexpr GemmSpecialization GemmSpec = GemmSpecialization::Default; static constexpr GemmSpecialization GemmSpec = GemmSpecialization::Default;
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false> template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
...@@ -359,141 +372,12 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -359,141 +372,12 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
batch)[I2]; batch)[I2];
} }
static constexpr index_t ClusterLengthMPerBlock = using NGCHWTransposeDescType =
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock::At(1); remove_cvref_t<decltype(conv_ngchw_to_nhwgc_transformer
static constexpr index_t ClusterLengthNPerBlock = .template MakeNGCHWTransposeDesc<NDimSpatial>({}, {}))>;
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock::At(3); using NHWGCTransposeDescType =
remove_cvref_t<decltype(conv_ngchw_to_nhwgc_transformer
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false> .template MakeNHWGCTransposeDesc<NDimSpatial>({}, {}))>;
static auto MakeInputTransposeDesc(std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_lengths,
std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_strides)
{
const index_t& G = g_n_c_wis_lengths[0];
const index_t& N = g_n_c_wis_lengths[1];
const index_t& C = g_n_c_wis_lengths[2];
const index_t& Hi = g_n_c_wis_lengths[3];
const index_t& Wi = g_n_c_wis_lengths[4];
const index_t& GStride = g_n_c_wis_strides[0];
const index_t& NStride = g_n_c_wis_strides[1];
const index_t& CStride = g_n_c_wis_strides[2];
const index_t& HiStride = g_n_c_wis_strides[3];
const index_t& WiStride = g_n_c_wis_strides[4];
const auto desc = make_naive_tensor_descriptor(
make_tuple(N, G, C, Hi, Wi), make_tuple(NStride, GStride, CStride, HiStride, WiStride));
const auto merged_desc =
transform_tensor_descriptor(desc,
make_tuple(make_merge_transform(make_tuple(N, G, C)),
make_merge_transform(make_tuple(Hi, Wi))),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return PadTensorDescriptor(
merged_desc,
make_tuple(MPerBlock / ClusterLengthMPerBlock, NPerBlock / ClusterLengthNPerBlock),
Sequence<true, true>{});
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
static auto MakeOutputTransposeDesc(std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_lengths,
std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_strides)
{
const index_t& G = g_n_c_wis_lengths[0];
const index_t& N = g_n_c_wis_lengths[1];
const index_t& C = g_n_c_wis_lengths[2];
const index_t& Hi = g_n_c_wis_lengths[3];
const index_t& Wi = g_n_c_wis_lengths[4];
const index_t& NStride = g_n_c_wis_strides[1];
const index_t HiStride = Wi * G * C;
const index_t WiStride = G * C;
const index_t GStride = C;
const index_t CStride = 1;
const auto desc = make_naive_tensor_descriptor(
make_tuple(N, G, C, Hi, Wi), make_tuple(NStride, GStride, CStride, HiStride, WiStride));
const auto merged_desc =
transform_tensor_descriptor(desc,
make_tuple(make_merge_transform(make_tuple(N, G, C)),
make_merge_transform(make_tuple(Hi, Wi))),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return PadTensorDescriptor(
merged_desc,
make_tuple(MPerBlock / ClusterLengthMPerBlock, NPerBlock / ClusterLengthNPerBlock),
Sequence<true, true>{});
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
static auto MakeInputTransposeDesc(std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_lengths,
std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_strides)
{
const index_t& G = g_n_c_wis_lengths[0];
const index_t& N = g_n_c_wis_lengths[1];
const index_t& C = g_n_c_wis_lengths[2];
const index_t& Di = g_n_c_wis_lengths[3];
const index_t& Hi = g_n_c_wis_lengths[4];
const index_t& Wi = g_n_c_wis_lengths[5];
const index_t& GStride = g_n_c_wis_strides[0];
const index_t& NStride = g_n_c_wis_strides[1];
const index_t& CStride = g_n_c_wis_strides[2];
const index_t& DiStride = g_n_c_wis_strides[3];
const index_t& HiStride = g_n_c_wis_strides[4];
const index_t& WiStride = g_n_c_wis_strides[5];
const auto desc = make_naive_tensor_descriptor(
make_tuple(N, G, C, Di, Hi, Wi),
make_tuple(NStride, GStride, CStride, DiStride, HiStride, WiStride));
const auto merged_desc =
transform_tensor_descriptor(desc,
make_tuple(make_merge_transform(make_tuple(N, G, C)),
make_merge_transform(make_tuple(Di, Hi, Wi))),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return PadTensorDescriptor(
merged_desc,
make_tuple(MPerBlock / ClusterLengthMPerBlock, NPerBlock / ClusterLengthNPerBlock),
Sequence<true, true>{});
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
static auto MakeOutputTransposeDesc(std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_lengths,
std::array<ck::index_t, NDimSpatial + 3> g_n_c_wis_strides)
{
const index_t& G = g_n_c_wis_lengths[0];
const index_t& N = g_n_c_wis_lengths[1];
const index_t& C = g_n_c_wis_lengths[2];
const index_t& Di = g_n_c_wis_lengths[3];
const index_t& Hi = g_n_c_wis_lengths[4];
const index_t& Wi = g_n_c_wis_lengths[5];
const index_t& NStride = g_n_c_wis_strides[1];
const index_t DiStride = Hi * Wi * G * C;
const index_t HiStride = Wi * G * C;
const index_t WiStride = G * C;
const index_t GStride = C;
const index_t CStride = 1;
const auto desc = make_naive_tensor_descriptor(
make_tuple(N, G, C, Di, Hi, Wi),
make_tuple(NStride, GStride, CStride, DiStride, HiStride, WiStride));
const auto merged_desc =
transform_tensor_descriptor(desc,
make_tuple(make_merge_transform(make_tuple(N, G, C)),
make_merge_transform(make_tuple(Di, Hi, Wi))),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return PadTensorDescriptor(
merged_desc,
make_tuple(MPerBlock / ClusterLengthMPerBlock, NPerBlock / ClusterLengthNPerBlock),
Sequence<true, true>{});
}
using InputTransposeDescType =
remove_cvref_t<decltype(MakeInputTransposeDesc<NDimSpatial>({}, {}))>;
using OutputTransposeDescType =
remove_cvref_t<decltype(MakeOutputTransposeDesc<NDimSpatial>({}, {}))>;
using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>()); using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
...@@ -572,8 +456,8 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -572,8 +456,8 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
I1>; I1>;
using GridwiseElementwiseTranspose = using GridwiseElementwiseTranspose =
GridwiseElementwise<Tuple<InputTransposeDescType>, GridwiseElementwise<Tuple<NGCHWTransposeDescType>,
Tuple<OutputTransposeDescType>, Tuple<NHWGCTransposeDescType>,
Tuple<const ADataType*>, Tuple<const ADataType*>,
Tuple<ADataType*>, Tuple<ADataType*>,
Block2TileMapElementwise, Block2TileMapElementwise,
...@@ -652,43 +536,11 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -652,43 +536,11 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
begin(output_spatial_lengths_)); begin(output_spatial_lengths_));
std::array<index_t, NDimSpatial + 3> b_g_n_c_wis_strides_transposed = std::array<index_t, NDimSpatial + 3> b_g_n_c_wis_strides_transposed =
b_g_n_c_wis_strides; conv_ngchw_to_nhwgc_transformer.TransposeStrides(b_g_n_c_wis_lengths,
b_g_n_c_wis_strides);
std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_strides_transposed = std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_strides_transposed =
a_g_n_k_wos_strides; conv_ngchw_to_nhwgc_transformer.TransposeStrides(a_g_n_k_wos_lengths,
a_g_n_k_wos_strides);
// NGKHW - transpose needed
if constexpr(is_NGCHW_GKYXC_NGKHW<InLayout, WeiLayout, OutLayout>() ||
is_NGCDHW_GKZYXC_NGKDHW<InLayout, WeiLayout, OutLayout>())
{
b_g_n_c_wis_strides_transposed[I0] = Conv_C_;
b_g_n_c_wis_strides_transposed[I2] = I1;
a_g_n_k_wos_strides_transposed[I0] = Conv_K_;
a_g_n_k_wos_strides_transposed[I2] = I1;
if constexpr(NDimSpatial == 2)
{
b_g_n_c_wis_strides_transposed[I3] =
input_spatial_lengths_[I1] * Conv_G_ * Conv_C_;
b_g_n_c_wis_strides_transposed[I4] = Conv_G_ * Conv_C_;
a_g_n_k_wos_strides_transposed[I3] =
output_spatial_lengths_[I1] * Conv_G_ * Conv_K_;
a_g_n_k_wos_strides_transposed[I4] = Conv_G_ * Conv_K_;
}
else if constexpr(NDimSpatial == 3)
{
b_g_n_c_wis_strides_transposed[I3] =
input_spatial_lengths_[I1] * input_spatial_lengths_[I2] * Conv_G_ * Conv_C_;
b_g_n_c_wis_strides_transposed[I4] =
input_spatial_lengths_[I2] * Conv_G_ * Conv_C_;
b_g_n_c_wis_strides_transposed[I5] = Conv_G_ * Conv_C_;
a_g_n_k_wos_strides_transposed[I3] = output_spatial_lengths_[I1] *
input_spatial_lengths_[I2] * Conv_G_ *
Conv_K_;
a_g_n_k_wos_strides_transposed[I4] =
input_spatial_lengths_[I2] * Conv_G_ * Conv_K_;
a_g_n_k_wos_strides_transposed[I5] = Conv_G_ * Conv_K_;
}
}
const auto descs = const auto descs =
conv_to_gemm_transformer_v2 conv_to_gemm_transformer_v2
...@@ -755,14 +607,18 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -755,14 +607,18 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
is_NGCDHW_GKZYXC_NGKDHW<InLayout, WeiLayout, OutLayout>()) is_NGCDHW_GKZYXC_NGKDHW<InLayout, WeiLayout, OutLayout>())
{ {
a_in_transpose_desc_ = a_in_transpose_desc_ =
MakeInputTransposeDesc<NDimSpatial>(a_g_n_k_wos_lengths, a_g_n_k_wos_strides); conv_ngchw_to_nhwgc_transformer.template MakeNGCHWTransposeDesc<NDimSpatial>(
a_g_n_k_wos_lengths, a_g_n_k_wos_strides);
a_out_transpose_desc_ = a_out_transpose_desc_ =
MakeOutputTransposeDesc<NDimSpatial>(a_g_n_k_wos_lengths, a_g_n_k_wos_strides); conv_ngchw_to_nhwgc_transformer.template MakeNHWGCTransposeDesc<NDimSpatial>(
a_g_n_k_wos_lengths, a_g_n_k_wos_strides);
b_in_transpose_desc_ = b_in_transpose_desc_ =
MakeInputTransposeDesc<NDimSpatial>(b_g_n_c_wis_lengths, b_g_n_c_wis_strides); conv_ngchw_to_nhwgc_transformer.template MakeNGCHWTransposeDesc<NDimSpatial>(
b_g_n_c_wis_lengths, b_g_n_c_wis_strides);
b_out_transpose_desc_ = b_out_transpose_desc_ =
MakeOutputTransposeDesc<NDimSpatial>(b_g_n_c_wis_lengths, b_g_n_c_wis_strides); conv_ngchw_to_nhwgc_transformer.template MakeNHWGCTransposeDesc<NDimSpatial>(
b_g_n_c_wis_lengths, b_g_n_c_wis_strides);
elementwise_block_2_ctile_map_transpose_a_ = Block2TileMapElementwise{ elementwise_block_2_ctile_map_transpose_a_ = Block2TileMapElementwise{
a_in_transpose_desc_.GetLength(I0), a_in_transpose_desc_.GetLength(I1)}; a_in_transpose_desc_.GetLength(I0), a_in_transpose_desc_.GetLength(I1)};
...@@ -816,8 +672,8 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -816,8 +672,8 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
Block2TileMapElementwise elementwise_block_2_ctile_map_transpose_a_, Block2TileMapElementwise elementwise_block_2_ctile_map_transpose_a_,
elementwise_block_2_ctile_map_transpose_b_; elementwise_block_2_ctile_map_transpose_b_;
InputTransposeDescType a_in_transpose_desc_, b_in_transpose_desc_; NGCHWTransposeDescType a_in_transpose_desc_, b_in_transpose_desc_;
OutputTransposeDescType a_out_transpose_desc_, b_out_transpose_desc_; NHWGCTransposeDescType a_out_transpose_desc_, b_out_transpose_desc_;
// for computing batch offset // for computing batch offset
ComputePtrOffsetOfStridedBatch<I1, I1, I0> compute_ptr_offset_of_batch_; ComputePtrOffsetOfStridedBatch<I1, I1, I0> compute_ptr_offset_of_batch_;
...@@ -1569,13 +1425,14 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -1569,13 +1425,14 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
(arg.GetWorkspaceETensorSizeBytes() + arg.GetWorkspaceATensorSizeBytes()) / (arg.GetWorkspaceETensorSizeBytes() + arg.GetWorkspaceATensorSizeBytes()) /
sizeof(BDataType); sizeof(BDataType);
// Different data type for A and B is not supported
auto kernel_transpose = kernel_elementwise_dual<GridwiseElementwiseTranspose, auto kernel_transpose = kernel_elementwise_dual<GridwiseElementwiseTranspose,
ck::Tuple<InputTransposeDescType>, ck::Tuple<NGCHWTransposeDescType>,
ck::Tuple<InputTransposeDescType>, ck::Tuple<NGCHWTransposeDescType>,
ck::Tuple<OutputTransposeDescType>, ck::Tuple<NHWGCTransposeDescType>,
ck::Tuple<OutputTransposeDescType>, ck::Tuple<NHWGCTransposeDescType>,
ck::Tuple<const ADataType*>, ck::Tuple<const ADataType*>,
ck::Tuple<BDataType*>, ck::Tuple<ADataType*>,
Block2TileMapElementwise, Block2TileMapElementwise,
Block2TileMapElementwise, Block2TileMapElementwise,
element_wise::PassThrough>; element_wise::PassThrough>;
......
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