Skip to content

GitLab

Commit e0041ad8 authored by Adam Osewski's avatar Adam Osewski
Browse files

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

parents 3239201e ac9e01e2
Hide whitespace changes
Inline Side-by-side
Showing with 989 additions and 56 deletions
.github/dependabot.yml 0 → 100644
View file @ e0041ad8
# To get started with Dependabot version updates, you'll need to specify which
# package ecosystems to update and where the package manifests are located.
# Please see the documentation for all configuration options:
# https://docs.github.com/github/administering-a-repository/configuration-options-for-dependency-updates
version: 2
updates:
- package-ecosystem: "pip" # See documentation for possible values
directory: "/docs/sphinx" # Location of package manifests
open-pull-requests-limit: 10
schedule:
interval: "daily"
.gitignore
View file @ e0041ad8
......@@ -47,3 +47,12 @@ build*
# GDB temporary files
.gdb_history
install.dir*
# documentation artifacts
_build/
_images/
_static/
_templates/
_toc.yml
docBin/
_doxygen/
.readthedocs.yaml 0 → 100644
View file @ e0041ad8
# Read the Docs configuration file
# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details
version: 2
build:
os: ubuntu-22.04
tools:
python: "3.8"
sphinx:
configuration: docs/conf.py
formats: [htmlzip, pdf, epub]
python:
install:
- requirements: docs/sphinx/requirements.txt
CHANGELOG.md 0 → 100644
View file @ e0041ad8
# Change Log for Composable Kernel
Full documentation for Composable Kernel is not yet available.
## CK 0.2.0 for ROCm 5.5.0
### Fixed
- Fixed a bug in 6-dimensional kernels (#555).
- Fixed grouped ConvBwdWeight test case failure (#524).
### Optimizations
- Improve proformance of normalization kernel
### Added
- Added support on NAVI3x.
- Added user tutorial (#563).
- Added more instances for irregular GEMM sizes (#560).
- Added inter-wave consumer-producer programming model for GEMM kernels (#310).
- Added multi-D GEMM client APIs (#534).
- Added multi-embeddings support (#542).
- Added Navi3x blockwise GEMM and real GEMM support (#541).
- Added Navi grouped ConvBwdWeight support (#505).
### Changed
- Changed ...
CMakeLists.txt
View file @ e0041ad8
......@@ -22,6 +22,7 @@ include(TargetFlags)
list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip)
option(USE_BITINT_EXTENSION_INT4, "Whether to enable clang's BitInt extension to provide int4 data type." OFF)
option(USE_OPT_NAVI3X, "Whether to enable LDS cumode and Wavefront32 mode for NAVI3X silicons." OFF)
if(USE_BITINT_EXTENSION_INT4)
add_compile_definitions(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
......@@ -29,6 +30,12 @@ if(USE_BITINT_EXTENSION_INT4)
message("CK compiled with USE_BITINT_EXTENSION_INT4 set to ${USE_BITINT_EXTENSION_INT4}")
endif()
if(USE_OPT_NAVI3X)
add_compile_options(-mcumode)
add_compile_options(-mno-wavefrontsize64)
message("CK compiled with USE_OPT_NAVI3X set to ${USE_OPT_NAVI3X}")
endif()
## Threads
set(THREADS_PREFER_PTHREAD_FLAG ON)
find_package(Threads REQUIRED)
......
Dockerfile
View file @ e0041ad8
FROM ubuntu:20.04
ARG ROCMVERSION=5.3
ARG compiler_version="release"
ARG ROCMVERSION=5.6
ARG compiler_version=""
ARG compiler_commit=""
RUN set -xe
ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
# Add rocm repository
RUN apt-get update
RUN apt-get install -y wget gnupg
RUN wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
RUN sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"
RUN apt-get install -y wget gnupg curl
RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.6"]; then \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"; \
else sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \
apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \
amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914 && \
DEBIAN_FRONTEND=noninteractive amdgpu-install -y --usecase=rocm ; \
fi
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 http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
# Install dependencies
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
apt-utils \
build-essential \
ccache \
cmake-data \
cmake \
curl \
git \
hip-rocclr \
jq \
......@@ -37,11 +43,13 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
python-dev \
python3-dev \
python3-pip \
sshpass \
software-properties-common \
rocm-dev \
rocm-device-libs \
rocm-cmake \
vim \
nano \
zlib1g-dev \
openssh-server \
clang-format-10 \
......@@ -49,6 +57,17 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
apt-get clean && \
rm -rf /var/lib/apt/lists/*
#Install latest version of cmake
RUN apt purge --auto-remove -y cmake
RUN apt update
RUN apt install -y software-properties-common lsb-release
RUN apt clean all
RUN wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | tee /etc/apt/trusted.gpg.d/kitware.gpg >/dev/null
RUN apt-add-repository "deb https://apt.kitware.com/ubuntu/ $(lsb_release -cs) main"
RUN apt install -y kitware-archive-keyring
RUN rm /etc/apt/trusted.gpg.d/kitware.gpg
RUN apt install -y cmake
# Setup ubsan environment to printstacktrace
RUN ln -s /usr/bin/llvm-symbolizer-3.8 /usr/local/bin/llvm-symbolizer
ENV UBSAN_OPTIONS=print_stacktrace=1
......@@ -60,7 +79,7 @@ RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
ARG PREFIX=/opt/rocm
# Install packages for processing the performance results
RUN pip3 install --upgrade pip
RUN pip3 install sqlalchemy
RUN pip3 install sqlalchemy==1.4.46
RUN pip3 install pymysql
RUN pip3 install pandas
RUN pip3 install setuptools-rust
......@@ -84,12 +103,7 @@ ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'"
RUN sh -c "echo compiler commit = '$compiler_commit'"
RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ]; then \
sed -i '/$HIP_CLANG_TARGET = chomp($HIP_CLANG_TARGET);/c\ chomp($HIP_CLANG_TARGET);' /opt/rocm/hip/bin/hipcc.pl && \
sed -i '/$HIP_CLANG_TARGET = chomp($HIP_CLANG_TARGET);/c\ chomp($HIP_CLANG_TARGET);' /opt/rocm/bin/hipcc.pl; \
fi
RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_commit" = "" ]; then \
RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \
git clone -b "$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 && \
......@@ -97,7 +111,7 @@ RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_com
else echo "using the release compiler"; \
fi
RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_commit" != "" ]; then \
RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
......
Jenkinsfile
View file @ e0041ad8
......@@ -14,12 +14,39 @@ def show_node_info() {
def runShell(String command){
def responseCode = sh returnStatus: true, script: "${command} > tmp.txt"
def output = readFile(file: "tmp.txt")
echo "tmp.txt contents: $output"
return (output != "")
}
def getDockerImageName(){
def img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
def img
if (params.ROCMVERSION != "5.6"){
if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
else{
if (params.COMPILER_COMMIT == ""){
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
}
else{
def commit = "${params.COMPILER_COMMIT}"[0..6]
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
}
}
}
else{
if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
else{
if (params.COMPILER_COMMIT == ""){
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
}
else{
def commit = "${params.COMPILER_COMMIT}"[0..6]
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
}
}
}
return img
}
......@@ -43,11 +70,11 @@ def build_compiler(){
compiler = '/opt/rocm/bin/hipcc'
}
else{
if (params.COMPILER_VERSION == "release"){
compiler = "/opt/rocm/llvm/bin/clang++"
if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
compiler = "/llvm-project/build/bin/clang++"
}
else{
compiler = "/llvm-project/build/bin/clang++"
compiler = "/opt/rocm/llvm/bin/clang++"
}
}
return compiler
......@@ -165,7 +192,7 @@ def cmake_build(Map conf=[:]){
if(conf.get("build_install","") == "true")
{
config_targets = 'install ' + config_targets
setup_args = ' -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install' + setup_args
setup_args = ' -DBUILD_DEV=On -DCMAKE_INSTALL_PREFIX=../install' + setup_args
} else{
setup_args = ' -DBUILD_DEV=On' + setup_args
}
......@@ -226,7 +253,7 @@ def buildHipClangJob(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION != "release"){
if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
......@@ -281,7 +308,7 @@ def runCKProfiler(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION != "release"){
if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
......@@ -414,12 +441,13 @@ def Build_CK(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION != "release"){
if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
def variant = env.STAGE_NAME
def retimage
def navi_node = 0
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
try {
......@@ -433,6 +461,9 @@ def Build_CK(Map conf=[:]){
else{
echo "GPU is OK"
}
if ( runShell('grep -n "gfx1030" clinfo.log') ){
navi_node = 1
}
}
}
}
......@@ -451,6 +482,9 @@ def Build_CK(Map conf=[:]){
else{
echo "GPU is OK"
}
if ( runShell('grep -n "gfx1030" clinfo.log') ){
navi_node = 1
}
}
}
}
......@@ -459,11 +493,22 @@ def Build_CK(Map conf=[:]){
{
cmake_build(conf)
dir("build"){
//run tests and examples
sh 'make -j check'
//we only need the ckProfiler to run the performance tests, so we pack and stash it
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash "ckProfiler.tar.gz"
//run tests and examples
sh 'make -j\$(( \$(nproc) / 2 )) check'
if (navi_node == 0 ){
//we only need the ckProfiler to run the performance tests, so we pack and stash it
//do not stash profiler on Navi nodes
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash "ckProfiler.tar.gz"
}
if (params.RUN_FULL_QA){
// build deb packages
sh 'make -j package'
archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb'
archiveArtifacts artifacts: 'composablekernel-tests_*.deb'
sh 'mv composablekernel-ckprofiler_*.deb ckprofiler_0.2.0_amd64.deb'
stash "ckprofiler_0.2.0_amd64.deb"
}
}
}
}
......@@ -530,6 +575,8 @@ def process_results(Map conf=[:]){
unstash "perf_splitK_gemm.log"
unstash "perf_onnx_gemm.log"
sh "./process_qa_data.sh"
unstash "ckprofiler_0.2.0_amd64.deb"
sh "sshpass -p ${env.ck_deb_pw} scp -o StrictHostKeyChecking=no ckprofiler_0.2.0_amd64.deb ${env.ck_deb_user}@${env.ck_deb_ip}:/var/www/html/composable_kernel/"
}
else{
// unstash perf files to master
......@@ -550,8 +597,9 @@ 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
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;COMPILER_VERSION=release
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open''' : ""
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true
0 21 * * * % ROCMVERSION=5.5;COMPILER_VERSION=release;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
pipeline {
agent none
......@@ -568,16 +616,16 @@ pipeline {
description: "Force building docker image (default: false), set to true if docker image needs to be updated.")
string(
name: 'ROCMVERSION',
defaultValue: '5.3',
description: 'Specify which ROCM version to use: 5.2.3, or 5.3 (default), etc.')
defaultValue: '5.6',
description: 'Specify which ROCM version to use: 5.6 (default).')
string(
name: 'COMPILER_VERSION',
defaultValue: 'release',
description: 'Specify which version of compiler to use: ck-9110, release (default), or amd-stg-open.')
defaultValue: '',
description: 'Specify which version of compiler to use: release, amd-stg-open, or leave blank (default).')
string(
name: 'COMPILER_COMMIT',
defaultValue: '',
description: 'Specify which commit of compiler branch to use: leave empty to use the latest commit (default), or use 8a82e4eb7ba28521ba9a9424a0315a8a16590424 commit of amd-stg-open branch.')
description: 'Specify which commit of compiler branch to use: leave blank to use the latest commit, or use 5541927df00eabd6a110180170eca7785d436ee3 (default) commit of amd-stg-open branch.')
string(
name: 'BUILD_COMPILER',
defaultValue: 'hipcc',
......@@ -639,12 +687,47 @@ pipeline {
{
parallel
{
stage("Build CK and run Tests")
stage("Build CK and run Tests on MI100/MI200/MI300")
{
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("gfx908 || gfx90a") }
environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 " """ }"
execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a" -DCMAKE_CXX_FLAGS="-O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
}
}
stage("Build CK and run Tests on MI100/MI200")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("gfx908 || gfx90a") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
}
}
stage("Build CK and run Tests on Navi")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("navi21") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1030" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......@@ -657,7 +740,7 @@ pipeline {
{
parallel
{
stage("Run ckProfiler: gfx908 or gfx90a")
stage("Run ckProfiler: gfx90*")
{
when {
beforeAgent true
......@@ -666,7 +749,7 @@ pipeline {
options { retry(2) }
agent{ label rocmnode("gfx908 || gfx90a")}
environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}"
setup_args = """ -DGPU_TARGETS="gfx908;gfx90a" -DBUILD_DEV=On """
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......@@ -681,7 +764,7 @@ pipeline {
options { retry(2) }
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}"
setup_args = """ -DGPU_TARGETS="gfx90a" -DBUILD_DEV=On """
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......
README.md
View file @ e0041ad8
# Composable Kernel
## Methodology
Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
CK utilizes two concepts to achieve performance portability and code maintainability:
* A tile-based programming model
* Algorithm complexity reduction for complex ML operators, using innovative technique we call "Tensor Coordinate Transformation".
![ALT](/doc/image/ck_component.png "CK Components")
![ALT](/docs/data/ck_component.png "CK Components")
## Code Structure
Current CK library are structured into 4 layers:
* "Templated Tile Operators" layer
* "Templated Kernel and Invoker" layer
* "Instantiated Kernel and Invoker" layer
* "Client API" layer
![ALT](/doc/image/ck_layer.png "CK Layers")
![ALT](/docs/data/ck_layer.png "CK Layers")
## Documentation
Run the steps below to build documentation locally.
```
cd docs
pip3 install -r sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
```
## Contributors
The list of developers and contributors is here: [Contributors](/CONTRIBUTORS.md)
## Citation
If you use CK, please use following citations:
* CK paper will be freely available on arXiv soon: [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???)
* [CITATION.cff](/CITATION.cff)
## License
CK is released under the MIT license. [License File](/LICENSE)
# Build CK
## Build docker image
```bash
DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile .
```
## Launch docker
```bash
docker run \
-it \
......@@ -50,10 +67,12 @@ ck:latest \
```
## Build CK
```bash
mkdir build && cd build
# Need to specify target ID, example below is for gfx908 and gfx90a
cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
......@@ -64,6 +83,7 @@ cmake
```
### Build examples and tests
```bash
make -j examples tests
make test
......@@ -73,21 +93,25 @@ Instructions for running each individual examples are under [example](/example)
## Build ckProfiler
```bash
make -j ckProfiler
```
Instructions for running ckProfiler are under [profiler](/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](/client_example)
## Caveat
### Kernel Timing and Verification
CK's own kernel timer will warn up kernel once, and then run it multiple times
to get average kernel time. For some kernels that use atomic add, this will cause
output buffer to be accumulated multiple times, causing verification failure.
......
client_example/01_gemm/gemm.cpp
View file @ e0041ad8
......@@ -83,7 +83,7 @@ int main(int argc, char* argv[])
[](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)
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
View file @ e0041ad8
......@@ -92,7 +92,7 @@ int main(int argc, char* argv[])
[](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)
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
View file @ e0041ad8
......@@ -88,7 +88,7 @@ int main(int argc, char* argv[])
[](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)
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
View file @ e0041ad8
......@@ -84,7 +84,7 @@ int main(int argc, char* argv[])
[](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)
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/03_gemm_layernorm/CMakeLists.txt
View file @ e0041ad8
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)
add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_operations)
client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp → client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
View file @ e0041ad8
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp"
......@@ -190,7 +190,7 @@ int main()
[](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)
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp 0 → 100644
View file @ e0041ad8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using F16 = ck::half_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd;
// DataType
using ADataType = F16;
using BDataType = F16;
using D0DataType = F16;
using D1DataType = F16;
using GammaDataType = F16;
using BetaDataType = F16;
using HDataType = F16;
// Layout
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using HLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddReluAdd;
using HElementOp = PassThrough;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
std::size_t mMemSize_;
};
int main(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideD0 = 0;
ck::index_t StrideD1 = N;
ck::index_t StrideH = N;
float epsilon = 1e-5;
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
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 d0_device_buf(sizeof(D0DataType) *
f_matrix_space_size(M, N, StrideD0, D0Layout{}));
SimpleDeviceMem d1_device_buf(sizeof(D1DataType) *
f_matrix_space_size(M, N, StrideD1, D1Layout{}));
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
SimpleDeviceMem h_device_buf(sizeof(HDataType) * f_matrix_space_size(M, N, StrideH, HLayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
HLayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
GammaDataType,
BetaDataType,
HDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddReluAdd,
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 cde_element_op = CDEElementOp{};
const auto h_element_op = HElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
h_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{StrideD0, StrideD1},
StrideH,
epsilon,
a_element_op,
b_element_op,
cde_element_op,
h_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace_dev(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
h_device_buf.SetZero();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
(sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
(sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
h_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{StrideD0, StrideD1},
StrideH,
epsilon,
a_element_op,
b_element_op,
cde_element_op,
h_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace_dev(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
h_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
\ No newline at end of file
client_example/04_contraction/CMakeLists.txt
View file @ e0041ad8
add_executable(client_contraction_scale contraction_scale.cpp)
target_link_libraries(client_contraction_scale PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_bilinear contraction_bilinear.cpp)
target_link_libraries(client_contraction_bilinear PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_operations)
add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_operations)
client_example/04_contraction/contraction_bilinear_fp64.cpp 0 → 100644
View file @ e0041ad8
// 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"
#include "ck/library/utility/numeric.hpp"
using F64 = double;
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 = F64;
using BDataType = F64;
using AccDataType = F64;
using CShuffleDataType = F64;
using DDataType = F64;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F64;
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[])
{
// kknn
#if 1
// 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};
// knnn
#elif 0
// 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{64, 1, 131072, 2048};
// 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};
// mknn
#elif 0
// 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{128, 1, 245760, 3840};
// 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};
// mnnn
#elif 0
// 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{128, 1, 245760, 3840};
// 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{64, 1, 131072, 2048};
// 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};
#endif
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 = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
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_g1m2n3k1_add_xdl_fp16.cpp 0 → 100644
View file @ e0041ad8
// 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_batched_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp"
#include "ck/library/utility/numeric.hpp"
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
static constexpr ck::index_t NumDimG = 1;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 3;
static constexpr ck::index_t NumDimK = 1;
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[])
{
ck::index_t G0 = 1;
ck::index_t M0 = 64;
ck::index_t M1 = 256;
ck::index_t N0 = 3;
ck::index_t N1 = 12;
ck::index_t N2 = 64;
ck::index_t K0 = 768;
// A[M0, M1, M2, K0]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, K0};
std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * K0, M1 * K0, K0, 1};
// B[N0, N1, N2, K0]
std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, N2, K0};
std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * N2 * K0, N1 * N2 * K0, N2 * K0, K0, 1};
// D[N0, M0, N1, M1, N2]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1 * N2, 0, 0, N1 * N2, N2, 1};
// E[N0 M0 N1 N2 M1]
std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
std::vector<ck::index_t> e_gs_ms_ns_strides{
M0 * M1 * N0 * N1 * N2, N1 * N2 * M1, 1, M0 * N1 * N2 * M1, M1 * N2, M1};
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_gs_ms_ks_lengths, a_gs_ms_ks_strides));
SimpleDeviceMem b_device_buf(sizeof(BDataType) *
f_tensor_space_size(b_gs_ns_ks_lengths, b_gs_ns_ks_strides));
SimpleDeviceMem d_device_buf(sizeof(DDataType) *
f_tensor_space_size(d_gs_ms_ns_lengths, d_gs_ms_ns_strides));
SimpleDeviceMem e_device_buf(sizeof(EDataType) *
f_tensor_space_size(e_gs_ms_ns_lengths, e_gs_ms_ns_strides));
using DeviceOp = ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
DsDataType,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Add>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
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_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b_gs_ns_ks_lengths,
b_gs_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
e_gs_ms_ns_lengths,
e_gs_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 = ck::accumulate_n<ck::index_t>(
e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
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;
}
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