ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-29 03:07:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8

Browse Source
This commit is contained in:
aska-0096
2024-12-13 09:18:50 +00:00
parent e8ca3daf4e 4e73177684
commit c8c016ddad
403 changed files with 24347 additions and 5555 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
.github/CODEOWNERS vendored
View File

@@ -1,8 +1,8 @@
* @junliume @illsilin @carlushuang @aosewski @poyenc @geyyer @bartekxk
* @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
# Documentation files
docs/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @aosewski @poyenc @geyyer @bartekxk
*.md @ROCm/rocm-documentation @junliume @illsilin @carlushuang @aosewski @poyenc @geyyer @bartekxk
*.rst @ROCm/rocm-documentation @junliume @illsilin @carlushuang @aosewski @poyenc @geyyer @bartekxk
.readthedocs.yaml @ROCm/rocm-documentation @junliume @illsilin @carlushuang @aosewski @poyenc @geyyer @bartekxk
docs/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
*.md @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
*.rst @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
.readthedocs.yaml @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
# Header directory for Doxygen documentation
library/include/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @aosewski @poyenc @geyyer @bartekxk
library/include/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca

11
CMakeLists.txt
View File

@@ -185,13 +185,22 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx9")
add_definitions(-DCK_USE_XDL)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx94")
message("Enabling FP8 gemms in ckProfiler")
message("Enabling FP8 gemms on native architectures")
add_definitions(-DCK_USE_GFX94)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
message("Enabling WMMA instances")
add_definitions(-DCK_USE_WMMA)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx12")
add_definitions(-DCK_USE_OCP_FP8)
set(CK_USE_OCP_FP8 "ON")
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_USE_FNUZ_FP8)
set(CK_USE_FNUZ_FP8 "ON")
endif()
option(CK_USE_FP8_ON_UNSUPPORTED_ARCH "Enable FP8 GEMM instances on older architectures" OFF)
if(CK_USE_FP8_ON_UNSUPPORTED_ARCH AND (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx908"))
add_definitions(-DCK_USE_FP8_ON_UNSUPPORTED_ARCH)

1
CONTRIBUTORS.md
View File

@@ -1,3 +1,4 @@
[Back to the main page](./README.md)
# Composable Kernel Developers and Contributors
This is the list of developers and contributors to Composable Kernel library

121
Dockerfile
View File

@@ -1,37 +1,28 @@
FROM ubuntu:20.04
FROM ubuntu:22.04
ARG DEBIAN_FRONTEND=noninteractive
ARG ROCMVERSION=6.2
ARG ROCMVERSION=6.3
ARG compiler_version=""
ARG compiler_commit=""
ARG CK_SCCACHE=""
RUN set -xe
ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
# Add rocm repository
RUN chmod 1777 /tmp
RUN apt-get update
RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN if [ "$ROCMVERSION" != "6.3" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.2.60200-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.2.60200-1_all.deb && \
# Add rocm repository
RUN set -xe && \
useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins && \
apt-get update && apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl && \
curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN if [ "$ROCMVERSION" != "6.4" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.3.60300-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.3.60300-1_all.deb && \
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] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$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_6.3-20.04-1_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_6.3-20.04-1_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 rel-20 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2074281; \
fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
RUN amdgpu-install -y --usecase=rocm --no-dkms
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" && \
amdgpu-install -y --usecase=rocm --no-dkms
## Sccache binary built from source for ROCm, only install if CK_SCCACHE is defined
ARG SCCACHE_REPO_URL=http://compute-artifactory.amd.com/artifactory/rocm-generic-experimental/rocm-sccache
@@ -57,6 +48,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
libnuma-dev \
libpthread-stubs0-dev \
llvm-amdgpu \
mpich \
net-tools \
pkg-config \
python \
@@ -76,68 +68,47 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
clang-format-12 \
kmod && \
apt-get clean && \
rm -rf /var/lib/apt/lists/*
rm -rf /var/lib/apt/lists/* && \
rm -rf amdgpu-install* && \
# Remove unnecessary rocm components that take a lot of space
apt-get remove -y rocblas rocfft rocsparse composablekernel-dev hipblaslt
# hipTensor requires rocm-llvm-dev for rocm versions > 6.0.1
RUN if [ "$ROCMVERSION" = "6.1" ]; then \
sh -c "apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated rocm-llvm-dev"; \
fi
# Update the cmake to version 3.27.5
RUN pip install --upgrade cmake==3.27.5
RUN pip install --upgrade cmake==3.27.5 && \
#Install latest ccache
RUN git clone https://github.com/ccache/ccache.git && \
cd ccache && mkdir build && cd build && cmake .. && make install
git clone https://github.com/ccache/ccache.git && \
cd ccache && mkdir build && cd build && cmake .. && make install && \
#Install ninja build tracing tools
RUN wget -qO /usr/local/bin/ninja.gz https://github.com/ninja-build/ninja/releases/latest/download/ninja-linux.zip
RUN gunzip /usr/local/bin/ninja.gz
RUN chmod a+x /usr/local/bin/ninja
RUN git clone https://github.com/nico/ninjatracing.git
cd / && \
wget -qO /usr/local/bin/ninja.gz https://github.com/ninja-build/ninja/releases/latest/download/ninja-linux.zip && \
gunzip /usr/local/bin/ninja.gz && \
chmod a+x /usr/local/bin/ninja && \
git clone https://github.com/nico/ninjatracing.git && \
#Install latest cppcheck
RUN git clone https://github.com/danmar/cppcheck.git && \
cd cppcheck && mkdir build && cd build && cmake .. && cmake --build .
WORKDIR /
# 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
git clone https://github.com/danmar/cppcheck.git && \
cd cppcheck && mkdir build && cd build && cmake .. && cmake --build . && \
cd / && \
# Install an init system
RUN wget https://github.com/Yelp/dumb-init/releases/download/v1.2.0/dumb-init_1.2.0_amd64.deb
RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
ARG PREFIX=/opt/rocm
wget https://github.com/Yelp/dumb-init/releases/download/v1.2.0/dumb-init_1.2.0_amd64.deb && \
dpkg -i dumb-init_*.deb && rm dumb-init_*.deb && \
# Install packages for processing the performance results
RUN pip3 install --upgrade pip
RUN pip3 install sqlalchemy==1.4.46
RUN pip3 install pymysql
RUN pip3 install pandas==2.0.3
RUN pip3 install setuptools-rust
RUN pip3 install sshtunnel==0.4.0
# Setup ubsan environment to printstacktrace
ENV UBSAN_OPTIONS=print_stacktrace=1
ENV LC_ALL=C.UTF-8
ENV LANG=C.UTF-8
RUN groupadd -f render
pip3 install --upgrade pip && \
pip3 install sqlalchemy==1.4.46 pymysql pandas==2.0.3 setuptools-rust sshtunnel==0.4.0 && \
# Add render group
groupadd -f render && \
# Install the new rocm-cmake version
RUN git clone -b master https://github.com/ROCm/rocm-cmake.git && \
cd rocm-cmake && mkdir build && cd build && \
cmake .. && cmake --build . && cmake --build . --target install
git clone -b master https://github.com/ROCm/rocm-cmake.git && \
cd rocm-cmake && mkdir build && cd build && \
cmake .. && cmake --build . && cmake --build . --target install
WORKDIR /
# Add alternative compilers, if necessary
ENV compiler_version=$compiler_version
ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'"
RUN sh -c "echo compiler commit = '$compiler_commit'"
RUN sh -c "echo compiler version = '$compiler_version'" && \
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" ] ) && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/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" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
@@ -145,16 +116,10 @@ RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd
else echo "using the release compiler"; \
fi
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline-open" ] ) && [ "$compiler_commit" != "" ]; then \
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline" ] ) && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
make -j 8 ; \
else echo "using the release compiler"; \
fi
#clean-up the deb package
RUN sh -c "rm -rf amdgpu-install*"
#ENV HIP_CLANG_PATH='/llvm-project/build/bin'
#RUN sh -c "echo HIP_CLANG_PATH = '$HIP_CLANG_PATH'"

26
Dockerfile.compiler Normal file
View File

@@ -0,0 +1,26 @@
ARG BASE_DOCKER="rocm/composable_kernel:ck_ub22.04_rocm6.3"
FROM $BASE_DOCKER
ARG compiler_version=""
ARG compiler_commit=""
# Add alternative compilers, if necessary
ENV compiler_version=$compiler_version
ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'" && \
sh -c "echo compiler commit = '$compiler_commit'"
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline" ] ) && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
cd llvm-project && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
make -j 16 ; \
else echo "using the release compiler"; \
fi
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline" ] ) && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
cd llvm-project && git checkout "$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" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
make -j 16 ; \
else echo "using the release compiler"; \
fi

368
Jenkinsfile vendored
View File

@@ -32,41 +32,43 @@ def runShell(String command){
return (output != "")
}
def getDockerImageName(){
def getBaseDockerImageName(){
def img
if (params.USE_CUSTOM_DOCKER != ""){
img = "${params.USE_CUSTOM_DOCKER}"
}
else{
if (params.ROCMVERSION != "6.3"){
if (params.COMPILER_VERSION == "") {
img = "${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}_${commit}"
}
}
def ROCM_numeric = "${params.ROCMVERSION}" as float
if ( ROCM_numeric < 6.4 ){
img = "${env.CK_DOCKERHUB}:ck_ub22.04_rocm${params.ROCMVERSION}"
}
else{
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub22.04_rocm${params.ROCMVERSION}"
}
}
return img
}
def getDockerImageName(){
def img
def base_name = getBaseDockerImageName()
if (params.USE_CUSTOM_DOCKER != ""){
img = "${params.USE_CUSTOM_DOCKER}"
}
else{
if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}"
img = "${base_name}"
}
else{
if (params.COMPILER_COMMIT == ""){
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
img = "${base_name}_${params.COMPILER_VERSION}"
}
else{
def commit = "${params.COMPILER_COMMIT}"[0..6]
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
img = "${base_name}_${params.COMPILER_VERSION}_${commit}"
}
}
}
}
return img
}
@@ -131,17 +133,21 @@ def buildDocker(install_prefix){
env.DOCKER_BUILDKIT=1
checkout scm
def image_name = getDockerImageName()
def base_image_name = getBaseDockerImageName()
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}' --build-arg DISABLE_CACHE='git rev-parse ${params.COMPILER_VERSION}' "
if(params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
dockerArgs = dockerArgs + " --no-cache "
def dockerArgs = "--build-arg PREFIX=${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}' "
if(params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline" || params.COMPILER_COMMIT != ""){
dockerArgs = dockerArgs + " --no-cache --build-arg BASE_DOCKER='${base_image_name}' -f Dockerfile.compiler . "
}
else{
dockerArgs = dockerArgs + " -f Dockerfile . "
}
echo "Build Args: ${dockerArgs}"
try{
if(params.BUILD_DOCKER){
//force building the new docker if that parameter is true
echo "Building image: ${image_name}"
retimage = docker.build("${image_name}", dockerArgs + ' .')
retimage = docker.build("${image_name}", dockerArgs)
withDockerRegistry([ credentialsId: "docker_test_cred", url: "" ]) {
retimage.push()
}
@@ -324,10 +330,8 @@ def cmake_build(Map conf=[:]){
try{
archiveArtifacts "perf_fmha_fwd_*.log"
archiveArtifacts "perf_fmha_bwd_*.log"
stash name: "perf_fmha_fwd_gfx942.log"
stash name: "perf_fmha_bwd_gfx942.log"
stash name: "perf_fmha_fwd_gfx90a.log"
stash name: "perf_fmha_bwd_gfx90a.log"
stash includes: "perf_fmha_**_gfx942.log", name: "perf_fmha_log_gfx942"
stash includes: "perf_fmha_**_gfx90a.log", name: "perf_fmha_log_gfx90a"
}
catch(Exception err){
echo "could not locate the requested artifacts: ${err.getMessage()}. will skip the stashing."
@@ -353,12 +357,12 @@ def buildHipClangJob(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="-u root --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${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}' "
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" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3')
@@ -373,7 +377,7 @@ def buildHipClangJob(Map conf=[:]){
gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 48, unit: 'HOURS')
timeout(time: 20, unit: 'HOURS')
{
cmake_build(conf)
}
@@ -402,128 +406,6 @@ def buildHipClangJobAndReboot(Map conf=[:]){
}
}
def runCKProfiler(Map conf=[:]){
show_node_info()
env.HSA_ENABLE_SDMA=0
checkout scm
def image = getDockerImageName()
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3')
def render_id = sh(returnStdout: true, script: 'getent group render | cut -d: -f3')
dockerOpts = dockerOpts + " --group-add=${video_id} --group-add=${render_id} "
echo "Docker flags: ${dockerOpts}"
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
try {
(retimage, image) = getDockerImage(conf)
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
sh 'rocminfo | tee rocminfo.log'
if ( !runShell('grep -n "gfx" rocminfo.log') ){
throw new Exception ("GPU not found")
}
else{
echo "GPU is OK"
}
}
}
}
catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
echo "The job was cancelled or aborted"
throw e
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 24, unit: 'HOURS')
{
sh """
rm -rf build
mkdir build
"""
dir("build"){
unstash 'ckProfiler.tar.gz'
sh 'tar -xvf ckProfiler.tar.gz'
}
dir("script"){
if (params.RUN_FULL_QA){
sh "./run_full_performance_tests.sh 0 QA_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
archiveArtifacts "perf_batched_gemm.log"
archiveArtifacts "perf_grouped_gemm.log"
archiveArtifacts "perf_grouped_conv_fwd.log"
archiveArtifacts "perf_grouped_conv_bwd_data.log"
archiveArtifacts "perf_grouped_conv_bwd_weight.log"
archiveArtifacts "perf_gemm_bilinear.log"
archiveArtifacts "perf_reduction.log"
archiveArtifacts "perf_splitK_gemm.log"
archiveArtifacts "perf_onnx_gemm.log"
archiveArtifacts "perf_mixed_gemm.log"
// stash perf files to master
stash name: "perf_gemm.log"
stash name: "perf_resnet50_N256.log"
stash name: "perf_resnet50_N4.log"
stash name: "perf_batched_gemm.log"
stash name: "perf_grouped_gemm.log"
stash name: "perf_grouped_conv_fwd.log"
stash name: "perf_grouped_conv_bwd_data.log"
stash name: "perf_grouped_conv_bwd_weight.log"
stash name: "perf_gemm_bilinear.log"
stash name: "perf_reduction.log"
stash name: "perf_splitK_gemm.log"
stash name: "perf_onnx_gemm.log"
stash name: "perf_mixed_gemm.log"
//we will process results on the master node
}
else{
sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
// stash perf files to master
stash name: "perf_gemm.log"
stash name: "perf_resnet50_N256.log"
stash name: "perf_resnet50_N4.log"
//we will process the results on the master node
}
}
}
}
}
return retimage
}
def runPerfTest(Map conf=[:]){
try{
runCKProfiler(conf)
}
catch(e){
echo "throwing error exception in performance tests"
echo 'Exception occurred: ' + e.toString()
throw e
}
finally{
if (!conf.get("no_reboot", false)) {
reboot()
}
}
}
def Build_CK(Map conf=[:]){
show_node_info()
@@ -544,12 +426,12 @@ def Build_CK(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="-u root --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
if(params.BUILD_LEGACY_OS){
@@ -567,7 +449,7 @@ def Build_CK(Map conf=[:]){
try {
(retimage, image) = getDockerImage(conf)
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
timeout(time: 2, unit: 'MINUTES'){
sh 'rocminfo | tee rocminfo.log'
if ( !runShell('grep -n "gfx" rocminfo.log') ){
throw new Exception ("GPU not found")
@@ -583,36 +465,95 @@ def Build_CK(Map conf=[:]){
throw e
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 24, unit: 'HOURS')
timeout(time: 20, unit: 'HOURS')
{
//check whether to run performance tests on this node
def do_perf_tests = 0
def arch_type = 0
sh 'rocminfo | tee rocminfo.log'
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx1201" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
do_perf_tests = 1
echo "Stash profiler and run performance tests"
if ( runShell('grep -n "gfx90a" rocminfo.log') ){
arch_type = 1
}
else if ( runShell('grep -n "gfx942" rocminfo.log') ) {
arch_type = 2
}
else if ( runShell('grep -n "gfx1030" rocminfo.log') ) {
arch_type = 3
}
else if ( runShell('grep -n "gfx1101" rocminfo.log') ) {
arch_type = 4
}
else if ( runShell('grep -n "gfx1201" rocminfo.log') ) {
arch_type = 5
}
cmake_build(conf)
dir("build"){
//run tests and examples
//sh 'make -j check'
if (params.RUN_PERFORMANCE_TESTS && do_perf_tests == 0 ){
//we only need the ckProfiler to run the performance tests, so we pack and stash it
//do not stash profiler on nodes where we don't need to run performance tests
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash name: "ckProfiler.tar.gz"
}
if (params.RUN_FULL_QA && do_perf_tests == 0 ){
// build deb packages for all gfx9 targets and prepare to export
if (params.RUN_FULL_QA && arch_type == 1 ){
// build deb packages for all gfx9 targets on gfx90a system and prepare to export
echo "Build ckProfiler package"
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 name: "ckprofiler_0.2.0_amd64.deb"
stash includes: "ckprofiler_0.2.0_amd64.deb", name: "ckprofiler_0.2.0_amd64.deb"
}
}
if (params.hipTensor_test && do_perf_tests == 0 ){
//build and test hipTensor
// run performance tests, stash the logs, results will be processed on the master node
dir("script"){
if (params.RUN_PERFORMANCE_TESTS){
if (params.RUN_FULL_QA && arch_type == 1){
// run full tests on gfx90a
echo "Run full performance tests"
sh "./run_full_performance_tests.sh 0 QA_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
archiveArtifacts "perf_batched_gemm.log"
archiveArtifacts "perf_grouped_gemm.log"
archiveArtifacts "perf_grouped_conv_fwd.log"
archiveArtifacts "perf_grouped_conv_bwd_data.log"
archiveArtifacts "perf_grouped_conv_bwd_weight.log"
archiveArtifacts "perf_gemm_bilinear.log"
archiveArtifacts "perf_reduction.log"
archiveArtifacts "perf_splitK_gemm.log"
archiveArtifacts "perf_onnx_gemm.log"
archiveArtifacts "perf_mixed_gemm.log"
stash includes: "perf_**.log", name: "perf_log"
}
else if ( arch_type == 1 ){
// run standard tests on gfx90a
echo "Run performance tests"
sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_onnx_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
stash includes: "perf_**.log", name: "perf_log"
}
// disable performance tests on gfx1030 for now.
//else if ( arch_type == 3){
// run basic tests on gfx1030
// echo "Run gemm performance tests"
// sh "./run_gemm_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME} gfx10"
// archiveArtifacts "perf_onnx_gemm_gfx10.log"
// stash includes: "perf_onnx_gemm_gfx10.log", name: "perf_log_gfx10"
//}
else if ( arch_type == 4){
// run basic tests on gfx11
echo "Run gemm performance tests"
sh "./run_gemm_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME} gfx11"
archiveArtifacts "perf_onnx_gemm_gfx11.log"
stash includes: "perf_onnx_gemm_gfx11.log", name: "perf_log_gfx11"
}
else if ( arch_type == 5 ){
// run basic tests on gfx12
echo "Run gemm performance tests"
sh "./run_gemm_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME} gfx12"
archiveArtifacts "perf_onnx_gemm_gfx12.log"
stash includes: "perf_onnx_gemm_gfx12.log", name: "perf_log_gfx12"
}
}
}
if (params.hipTensor_test && arch_type == 1 ){
// build and test hipTensor on gfx90a node
sh """#!/bin/bash
rm -rf "${params.hipTensor_branch}".zip
rm -rf hipTensor-"${params.hipTensor_branch}"
@@ -679,15 +620,13 @@ def process_results(Map conf=[:]){
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 1, unit: 'HOURS'){
timeout(time: 15, unit: 'MINUTES'){
try{
dir("script"){
if (params.RUN_CK_TILE_FMHA_TESTS){
try{
unstash "perf_fmha_fwd_gfx942.log"
unstash "perf_fmha_bwd_gfx942.log"
unstash "perf_fmha_fwd_gfx90a.log"
unstash "perf_fmha_bwd_gfx90a.log"
unstash "perf_fmha_log_gfx942"
unstash "perf_fmha_log_gfx90a"
}
catch(Exception err){
echo "could not locate the FMHA performance logs: ${err.getMessage()}."
@@ -697,26 +636,26 @@ def process_results(Map conf=[:]){
// unstash perf files to master
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/"
unstash "perf_gemm.log"
unstash "perf_resnet50_N256.log"
unstash "perf_resnet50_N4.log"
unstash "perf_batched_gemm.log"
unstash "perf_grouped_gemm.log"
unstash "perf_grouped_conv_fwd.log"
unstash "perf_grouped_conv_bwd_data.log"
unstash "perf_grouped_conv_bwd_weight.log"
unstash "perf_gemm_bilinear.log"
unstash "perf_reduction.log"
unstash "perf_splitK_gemm.log"
unstash "perf_onnx_gemm.log"
unstash "perf_mixed_gemm.log"
unstash "perf_log"
try{
unstash "perf_log_gfx11"
unstash "perf_log_gfx12"
}
catch(Exception err){
echo "could not locate the GEMM gfx11/gfx12 performance logs: ${err.getMessage()}."
}
sh "./process_qa_data.sh"
}
else{
// unstash perf files to master
unstash "perf_gemm.log"
unstash "perf_resnet50_N256.log"
unstash "perf_resnet50_N4.log"
unstash "perf_log"
try{
unstash "perf_log_gfx11"
unstash "perf_log_gfx12"
}
catch(Exception err){
echo "could not locate the GEMM gfx11/gfx12 performance logs: ${err.getMessage()}."
}
sh "./process_perf_data.sh"
}
}
@@ -734,10 +673,10 @@ 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;ROCMVERSION=6.2;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true;RUN_CODEGEN_TESTS=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
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.3;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.3;hipTensor_test=true;RUN_CODEGEN_TESTS=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
0 13 * * * % BUILD_LEGACY_OS=true''' : ""
@@ -760,12 +699,12 @@ pipeline {
description: 'If you want to use a custom docker image, please specify it here (default: leave blank).')
string(
name: 'ROCMVERSION',
defaultValue: '6.2',
description: 'Specify which ROCM version to use: 6.2 (default).')
defaultValue: '6.3',
description: 'Specify which ROCM version to use: 6.3 (default).')
string(
name: 'COMPILER_VERSION',
defaultValue: '',
description: 'Specify which version of compiler to use: release, amd-staging, amd-mainline-open, or leave blank (default).')
description: 'Specify which version of compiler to use: release, amd-staging, amd-mainline, or leave blank (default).')
string(
name: 'COMPILER_COMMIT',
defaultValue: '',
@@ -824,8 +763,8 @@ pipeline {
description: "Test building instances for various architectures simultaneously (default: OFF)")
booleanParam(
name: "BUILD_GFX12",
defaultValue: false,
description: "Build CK and run tests on gfx12 (default: OFF)")
defaultValue: true,
description: "Build CK and run tests on gfx12 (default: ON)")
booleanParam(
name: "NINJA_BUILD_TRACE",
defaultValue: false,
@@ -1235,29 +1174,6 @@ pipeline {
}
}
}
stage("Performance Tests")
{
parallel
{
stage("Run ckProfiler: gfx90a")
{
when {
beforeAgent true
expression { params.RUN_PERFORMANCE_TESTS.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
}
options { retry(1) }
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "NO_CK_BUILD"
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
cleanWs()
}
}
}
}
stage("Process Performance Test Results")
{
parallel

37
README.md
View File

@@ -26,23 +26,15 @@ The current CK library is structured into four layers:
## General information
To build our documentation locally, use the following code:
``` bash
cd docs
pip3 install -r sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
```
You can find a list of our developers and contributors on our [Contributors](/CONTRIBUTORS.md) page.
```note
If you use CK, cite us as follows:
* [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???):
This paper will be available on arXiv soon.
* [CITATION.cff](/CITATION.cff)
```
* [CK supported operations](include/ck/README.md)
* [CK Tile supported operations](include/ck_tile/README.md)
* [CK wrapper](client_example/25_wrapper/README.md)
* [CK codegen](codegen/README.md)
* [CK profiler](profiler/README.md)
* [Examples (Custom use of CK supported operations)](example/README.md)
* [Client examples (Use of CK supported operations with instance factory)](client_example/README.md)
* [Terminology](/TERMINOLOGY.md)
* [Contributors](/CONTRIBUTORS.md)
CK is released under the **[MIT license](/LICENSE)**.
@@ -137,6 +129,14 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa
You can find instructions for running ckProfiler in [profiler](/profiler).
* Build our documentation locally:
``` bash
cd docs
pip3 install -r sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
```
Note the `-j` option for building with multiple threads in parallel, which speeds up the build significantly.
However, `-j` launches unlimited number of threads, which can cause the build to run out of memory and
crash. On average, you should expect each thread to use ~2Gb of RAM.
@@ -154,8 +154,7 @@ Additional cmake flags can be used to significantly speed-up the build:
other platforms have faster instances, such as `xdl` or `wmma`, available.
* `CK_USE_FP8_ON_UNSUPPORTED_ARCH` (default is OFF) must be set to ON in order to build instances,
such as `gemm_universal` and `gemm_multiply_multiply` for fp8 data type for GPU targets which do not
have native support for fp8 data type, such as gfx908 or gfx90a. These instances are useful on
such as `gemm_universal`, `gemm_universal_streamk` and `gemm_multiply_multiply` for fp8 data type for GPU targets which do not have native support for fp8 data type, such as gfx908 or gfx90a. These instances are useful on
architectures like the MI100/MI200 for the functional support only.
## Using sccache for building

2
TERMINOLOGY.md Normal file
View File

@@ -0,0 +1,2 @@
[Back to the main page](./README.md)
# Composable Kernel terminology

4
client_example/24_grouped_conv_activation/CMakeLists.txt
View File

@@ -54,7 +54,7 @@ target_link_libraries(client_conv3d_fwd_convscale_relu_amax_fp8
PRIVATE composable_kernel::device_conv_operations
composable_kernel::device_other_operations
composable_kernel::device_reduction_operations
utility)
composable_kernel::utility)
# Fwd convscale + AMAX
add_executable(client_conv3d_fwd_convscale_amax_fp8
grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_amax_fp8.cpp)
@@ -62,7 +62,7 @@ target_link_libraries(client_conv3d_fwd_convscale_amax_fp8
PRIVATE composable_kernel::device_conv_operations
composable_kernel::device_other_operations
composable_kernel::device_reduction_operations
utility)
composable_kernel::utility)
# Fwd convscale
add_executable(client_conv3d_fwd_convscale_fp8
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8.cpp)

11
client_example/25_wrapper/README.md
View File

@@ -1,14 +1,9 @@
[Back to the main page](../../README.md)
# Composable Kernel wrapper GEMM tutorial
This tutorial demonstrates how to implement matrix multiplication using Composable Kernel (CK)
wrapper. We present the base version of GEMM without most of the available optimizations; however,
it's worth noting that CK has kernels with different optimizations.
This tutorial demonstrates how to implement matrix multiplication using Composable Kernel (CK) wrapper. We present the base version of GEMM without most of the available optimizations; however, it's worth noting that CK has kernels with different optimizations.
To implement these optimizations, you can use the CK wrapper or directly use available instances in
CK. You can also refer to the
[optimized GEMM example](https://github.com/ROCm/composable_kernel/blob/develop/client_example/25_wrapper/wrapper_optimized_gemm.cpp),
that uses CK wrapper based on the
[`gridwise_gemm_xdlops_v2r3`](https://github.com/ROCm/composable_kernel/blob/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp) implementation.
To implement these optimizations, you can use the CK wrapper or directly use available instances in CK. You can also refer to the [optimized GEMM example](https://github.com/ROCm/composable_kernel/blob/develop/client_example/25_wrapper/wrapper_optimized_gemm.cpp), that uses CK wrapper based on the [`gridwise_gemm_xdlops_v2r3`](https://github.com/ROCm/composable_kernel/blob/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp) implementation.
The kernel definition should look similar to:

2
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp
View File

@@ -121,7 +121,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
constexpr ck::index_t NumDTensor = 2;
using GroupedGemmKernelArgument =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
ck::tensor_operation::device::GroupedGemmKernelArgument<NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);

2
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_xdl_bf16_i8.cpp
View File

@@ -120,7 +120,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
constexpr ck::index_t NumDTensor = 1;
using GroupedGemmKernelArgument =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
ck::tensor_operation::device::GroupedGemmKernelArgument<NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);

10
client_example/CMakeLists.txt
View File

@@ -56,13 +56,21 @@ if (GPU_TARGETS)
add_definitions(-DCK_USE_WMMA)
set(CK_USE_WMMA "ON")
endif()
if (GPU_TARGETS MATCHES "gfx12")
add_definitions(-DCK_USE_OCP_FP8)
set(CK_USE_OCP_FP8 "ON")
endif()
if (GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_USE_FNUZ_FP8)
set(CK_USE_FNUZ_FP8 "ON")
endif()
else()
add_definitions(-DCK_USE_WMMA -DCK_USE_XDL)
set(CK_USE_XDL "ON")
set(CK_USE_WMMA "ON")
endif()
find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations device_reduction_operations)
find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations device_reduction_operations utility)
if(GPU_TARGETS MATCHES "gfx9")
find_package(composable_kernel COMPONENTS device_contraction_operations)
endif()

2
client_example/README.md
View File

@@ -1,3 +1,5 @@
[Back to the main page](../README.md)
# Composable Kernel client examples
##
Client application links to CK library, and therefore CK library needs to be installed before building client applications.

1
codegen/CMakeLists.txt
View File

@@ -7,6 +7,7 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/..)
configure_file(${CK_ROOT}/include/ck/config.h.in ${CK_ROOT}/include/ck/config.h)
find_package(ROCM)
include(ROCMInstallTargets)

2
codegen/README.md Normal file
View File

@@ -0,0 +1,2 @@
[Back to the main page](../README.md)
# Composable Kernel codegen

1
codegen/test/rtc/include/rtc/hip.hpp
View File

@@ -4,6 +4,7 @@
#include <hip/hip_runtime_api.h>
#include <memory>
#include <string>
#include <stdexcept>
namespace rtc {

2
docs/sphinx/requirements.in
View File

@@ -1,2 +1,2 @@
rocm-docs-core==1.8.5
rocm-docs-core==1.11.0
sphinxcontrib-bibtex==2.6.3

2
docs/sphinx/requirements.txt
View File

@@ -103,7 +103,7 @@ requests==2.32.3
# via
# pygithub
# sphinx
rocm-docs-core==1.8.5
rocm-docs-core==1.11.0
# via -r requirements.in
six==1.16.0
# via pybtex

3
example/01_gemm/CMakeLists.txt
View File

@@ -77,6 +77,9 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
add_example_executable(example_gemm_xdl_fp8_bf8 gemm_xdl_fp8_bf8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
add_example_executable(example_gemm_xdl_fp8_streamk_v3 gemm_xdl_fp8_streamk_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_streamk_v3)
add_example_executable(example_gemm_xdl_fp16_fp8 gemm_xdl_fp16_fp8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)

4
example/01_gemm/common.hpp
View File

@@ -44,7 +44,7 @@ struct ProblemSizeStreamK final
ck::index_t StrideB = -1;
ck::index_t StrideC = -1;
ck::index_t NumSKBlocks = -1;
ck::index_t NumSKBlocks = -1; // number of stream-k blocks
};
struct ProblemSizeStreamK_universal final
{
@@ -76,7 +76,7 @@ struct ProblemSizeSplitK final
struct ExecutionConfig final
{
// 0 - no verification, 1 - CPU, 2 - GPU, 3 - CPU + GPU
int do_verification = 3;
int do_verification = 1;
int init_method = 2;
bool time_kernel = false;
};

13
example/01_gemm/gemm_xdl_fp16_streamk_v3.cpp
View File

@@ -8,7 +8,7 @@
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CShuffleDataType = float;
using CDataType = ck::half_t;
using ALayout = Row;
@@ -43,6 +43,17 @@ using DeviceGemmV2_Streamk_Instance =
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example_streamk_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_universal_streamk_example(argc, argv); }

58
example/01_gemm/gemm_xdl_fp8_streamk_v3.cpp Executable file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::f8_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CDataType = ck::half_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmV2_Streamk_Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_Streamk_V3<
ALayout, BLayout, CLayout,
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault,
256,
128, 256,
128, 16, 16,
16, 16,
4, 8,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 1,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 1,
1, 2, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ck::f8_t>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example_streamk_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_universal_streamk_example(argc, argv); }

4
example/01_gemm/run_gemm_example.inc
View File

@@ -143,8 +143,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
switch(config.init_method)
{
case 0:
ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
ck::utils::FillConstant<ADataType>{ck::type_convert<ADataType>(1.f)}(a_m_k);
ck::utils::FillConstant<BDataType>{ck::type_convert<BDataType>(1.f)}(b_k_n);
break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);

40
example/01_gemm/run_gemm_example_streamk_v2.inc Normal file → Executable file
View File

@@ -176,6 +176,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
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_ref_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
@@ -196,6 +197,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_ref_buf(sizeof(CDataType) *
c_m_n_device_ref_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
@@ -240,6 +243,13 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
return true;
}
std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
if(workspace_size != 0)
{
workspace.Realloc(workspace_size);
gemm.SetWorkSpacePointer(&argument, workspace.GetDeviceBuffer());
}
bool pass = true;
if((config.do_verification == 1) || (config.do_verification == 3))
{
@@ -271,6 +281,36 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
#endif
}
if((config.do_verification == 2) || (config.do_verification == 3))
{
// GPU verification
auto ref_gemm_gpu = ReferenceGemmInstanceGPU{};
auto ref_invoker_gpu = ref_gemm_gpu.MakeInvoker();
auto ref_argument_gpu = ref_gemm_gpu.MakeArgument(
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_ref_buf.GetDeviceBuffer()),
M,
N,
K,
a_element_op,
b_element_op,
c_element_op);
std::cout << "Running verification on GPU." << std::endl;
ref_invoker_gpu.Run(ref_argument_gpu, StreamConfig{});
c_m_n_device_ref_buf.FromDevice(c_m_n_device_ref_result.mData.data());
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_device_ref_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
}
if(config.time_kernel)
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});

2
example/01_gemm/run_gemm_example_v2.inc
View File

@@ -261,7 +261,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
if(config.time_kernel)
{
ave_time =
invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 0, 5, 10, true, 4});
invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 0, 50, 100, true, 4});
std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype =

12
example/15_grouped_gemm/grouped_gemm_multiple_d_splitk_xdl_fp16.cpp
View File

@@ -186,15 +186,15 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
for(int j = 0; j < NumDMatrices; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
for(int j = 0; j < NumDMatrices; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<DDataType, 0>{});
}
}
}
@@ -246,7 +246,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
// do GEMM
auto argument = gemm.MakeArgument(
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, cde_element_op);
gemm.SetKBatchSize(argument, config.k_batch);
gemm.SetKBatchSize(&argument, config.k_batch);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
@@ -257,7 +257,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
gemm.SetDeviceKernelArgs(&argument, gemm_arg_dev_mem.GetDeviceBuffer());
invoker.Run(argument, StreamConfig{nullptr, false, 1});

10
example/15_grouped_gemm/grouped_gemm_multiple_d_xdl_fp16.cpp
View File

@@ -91,7 +91,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
{
auto group_count = problem_size.group_count;
using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDs>;
using KernelArguments = ck::tensor_operation::device::GroupedGemmKernelArgument<NumDs>;
using GemmDesc = ck::tensor_operation::device::GemmDesc;
// GEMM shape
@@ -190,15 +190,15 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<DDataType, 0>{});
}
}
}

10
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -167,11 +167,11 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
d0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
d0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<D0DataType, 1>{});
}
using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<1>;
@@ -254,7 +254,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
gemm.GetDeviceKernelArgSize(&argument),
hipMemcpyHostToDevice));
gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
gemm.SetDeviceKernelArgs(&argument, gemm_kernel_args_dev.GetDeviceBuffer());
gemm.SetKBatch(argument, config.k_batch);
invoker.Run(argument, StreamConfig{nullptr, false});

8
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -157,8 +157,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
}
@@ -239,7 +239,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
"not support this GEMM problem");
}
gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
gemm.SetDeviceKernelArgs(&argument, gemm_arg_dev_mem.GetDeviceBuffer());
gemm.SetKBatch(argument, config.k_batch);
invoker.Run(argument, StreamConfig{nullptr, false});

8
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16_fp8.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -158,8 +158,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
}
@@ -240,7 +240,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
"not support this GEMM problem");
}
gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
gemm.SetDeviceKernelArgs(&argument, gemm_arg_dev_mem.GetDeviceBuffer());
gemm.SetKBatch(argument, config.k_batch);
invoker.Run(argument, StreamConfig{nullptr, false});

25
example/15_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -1,3 +1,6 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
struct ProblemSize final
@@ -124,8 +127,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
}
@@ -168,9 +171,23 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
auto argument = gemm.MakeArgument(
p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
std::size_t kargs_size = gemm.GetDeviceKernelArgSize(&argument);
gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
DeviceMem gemm_workspace, gemm_kargs;
// The following is necessary since TwoStage kernel is using additional memory both
// for Workspace and kernel arguments.
if(kargs_size > 0)
{
gemm_kargs.Realloc(kargs_size);
gemm.SetDeviceKernelArgs(&argument, gemm_kargs.GetDeviceBuffer());
}
if(workspace_size > 0 && workspace_size != kargs_size)
{
gemm_workspace.Realloc(workspace_size);
gemm.SetWorkSpacePointer(&argument, gemm_workspace.GetDeviceBuffer());
}
if(!gemm.IsSupportedArgument(argument))
{

6
example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -175,8 +175,8 @@ int main(int argc, char* argv[])
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
c0_n_bias.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5});

4
example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -150,7 +150,7 @@ bool run_batched_gemm_gemm_example(int argc, char* argv[])
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -157,7 +157,7 @@ int run(int argc, char* argv[])
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -118,7 +118,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute_wmma.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -153,7 +153,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_cross_attention_wmma.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -178,7 +178,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -152,7 +152,7 @@ int run(int argc, char* argv[])
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_grouped_query_attention_forward_wmma.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -156,7 +156,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_multi_query_attention_forward_wmma.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -156,7 +156,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

4
example/32_batched_gemm_scale_softmax_gemm/run_self_attention_wmma.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
@@ -173,7 +173,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}

7
example/35_splitK_gemm/run_splitK_gemm_example.inc
View File

@@ -1,3 +1,6 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
struct ProblemSize final
@@ -66,8 +69,8 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());

2
example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
View File

@@ -377,7 +377,7 @@ int main(int argc, char* argv[])
break;
default:
a0_g_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
d00_g_m_n.GenerateTensorValue(GeneratorTensor_1<D00DataType>{1});
d01_g_m_n.GenerateTensorValue(GeneratorTensor_1<D01DataType>{1});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});

4
example/38_grouped_conv_bwd_data_multiple_d/common.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -41,7 +41,7 @@ struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
bool time_kernel = false;
};
#define DefaultConvParams \

4
example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute_xdl.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
@@ -248,7 +248,7 @@ int main(int argc, char* argv[])
d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});

8
example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -194,9 +194,9 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b1_tensors[i].GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
default:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<A0DataType, 0>{});
b0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<B1DataType, 1>{});
}
d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});

6
example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp
View File

@@ -184,9 +184,9 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
break;
default:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
a1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<A0DataType, 0>{});
a1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<A1DataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
}
d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});

9
example/62_convnd_activ/convscale/convnd_fwd_convscale_common.hpp
View File

@@ -172,12 +172,13 @@ bool run_grouped_conv_fwd(bool do_verification,
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
// values generated: -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 6});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());

3
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp
View File

@@ -205,7 +205,6 @@ int main(int argc, char* argv[])
a1_device_buf.ToDevice(a1_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
@@ -253,8 +252,6 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});

7
example/CMakeLists.txt
View File

@@ -54,6 +54,13 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any DPP examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dpp")
message("removing dpp example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any XDL examples if gfx9 targets are not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT EX_TARGETS MATCHES "gfx9" AND source MATCHES "_xdl")

2
example/README.md Normal file
View File

@@ -0,0 +1,2 @@
[Back to the main page](../README.md)
# Composable Kernel examples

15
example/ck_tile/01_fmha/codegen/cpp_symbol_map.py
View File

@@ -2,10 +2,17 @@
# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
# generate kernel instances to speed up compilation
DTYPE_MAP = {
"fp16": "ck_tile::fp16_t",
"bf16": "ck_tile::bf16_t",
"fp8" : "ck_tile::fp8_t"
FWD_DTYPE_MAP = {
"fp16" : "FmhaFwdFp16",
"bf16" : "FmhaFwdBf16",
"fp8" : "FmhaFwdFp8",
"fp8fp16": "FmhaFwdFp8Fp16",
"fp8bf16": "FmhaFwdFp8Bf16"
}
BWD_DTYPE_MAP = {
"fp16": "FmhaBwdFp16",
"bf16": "FmhaBwdBf16"
}
MASK_IMPL = {

14
example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py
View File

@@ -283,7 +283,7 @@ class FmhaBwdApiPool:
inners = inners + FMHA_BWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_pipeline_enum=BWD_DQDKDV_PIPELINE_ENUM_MAP[trait.pipeline],
F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_mask=get_mask_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias],
F_bias=BIAS_MAP[trait.bias], F_dbias=BOOL_MAP[trait.dbias], F_dropout_check=DROPOUT_CHECK_MAP[trait.dropout], F_dropout=DROPOUT_MAP[trait.dropout],
F_scheck=trait.scheck(spad1=spad1), F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_hdim=hdim, F_dtype=DTYPE_MAP[dtype],
F_scheck=trait.scheck(spad1=spad1), F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_hdim=hdim, F_dtype=BWD_DTYPE_MAP[dtype],
F_spad0=BOOL_MAP[trait.spad], F_spad1=BOOL_MAP[spad1], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_deterministic=BOOL_MAP[trait.deterministic])
@@ -360,7 +360,7 @@ class FmhaBwdDQDKDVKernel:
FMHA_BWD_DQ_DK_DV_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = BWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0,
@@ -469,7 +469,7 @@ def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
gen = list()
api_pool = FmhaBwdApiPool(mask_impl)
for dtype in DTYPE_MAP.keys():
for dtype in BWD_DTYPE_MAP.keys():
d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
if d == None:
continue
@@ -585,7 +585,7 @@ class FmhaBwdOGradDotOKernel:
FMHA_BWD_DOT_DO_O_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = BWD_DTYPE_MAP[self.F_dtype],
F_spad = BOOL_MAP[self.F_spad],
F_dvpad = BOOL_MAP[self.F_dvpad],
F_mode = MODE_MAP[self.F_mode],
@@ -616,7 +616,7 @@ def get_bwd_dot_do_o_blobs() -> List[FmhaBwdOGradDotOKernel]:
gen = list()
for dtype in DTYPE_MAP.keys():
for dtype in BWD_DTYPE_MAP.keys():
d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
if d == None:
continue
@@ -716,7 +716,7 @@ class FmhaBwdConvertQGradKernel:
FMHA_BWD_CONVERT_DQ_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = BWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_bm0,
F_bn0 = self.F_bn0,
F_spad = BOOL_MAP[self.F_spad],
@@ -751,7 +751,7 @@ def get_bwd_convert_dq_blobs() -> List[FmhaBwdConvertQGradKernel]:
gen = list()
for dtype in DTYPE_MAP.keys():
for dtype in BWD_DTYPE_MAP.keys():
d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
if d == None:
continue

11
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View File

@@ -282,7 +282,7 @@ class FmhaFwdApiPool:
F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
@@ -301,7 +301,7 @@ class FmhaFwdTileSize:
F_bk1 : int # tile size along kv gemm unroll
F_bk0max : int # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
F_rm0 : int # number of warps for gemm0 along q seqlen
F_rn0 : int # number of warps for gemm0 along k seqlen
F_rn0 : int # number of warps for gemm0 along k seqlen
F_rk0 : int # number of warps for gemm0 along head dim q (not used)
F_rm1 : int # number of warps for gemm1 along q seqlen
F_rn1 : int # number of warps for gemm1 along head dim v
@@ -339,7 +339,7 @@ class FmhaFwdKernel:
FMHA_FWD_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0,
@@ -462,6 +462,9 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
# no need lse/dropout kernels
for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', 'f', squant, mask))
elif dtype in ['fp8fp16', 'fp8bf16']:
# TODO
None
else:
assert False
return pipelines
@@ -469,7 +472,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
gen = list()
api_pool = FmhaFwdApiPool(mask_impl)
for dtype in DTYPE_MAP.keys():
for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_tile_dict_from_dtype(dtype)
if d == None:
continue

9
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_appendkv.py
View File

@@ -181,7 +181,7 @@ class FmhaFwdAppendKVApiPool:
inners = inners + FMHA_FWD_APPENDKV_API_INNER_DISPATCH.format(F_if=if_k, F_vlayout=LAYOUT_MAP[trait.vlayout],
F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_rope_check=ROPE_CHECK_MAP[trait.rope],
F_pagedkv=BOOL_MAP[trait.pagedkv], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_rope=ROPE_MAP[trait.rope], F_bs=trait.bs, F_bsk=trait.bsk, F_bd=trait.bd, F_bdv=trait.bdv, F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
F_rope=ROPE_MAP[trait.rope], F_bs=trait.bs, F_bsk=trait.bsk, F_bd=trait.bd, F_bdv=trait.bdv, F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
@@ -216,7 +216,7 @@ class FmhaFwdAppendKVKernel:
FMHA_FWD_APPENDKV_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bs = self.F_tile.F_bs,
F_bsk = self.F_tile.F_bsk,
F_bd = self.F_tile.F_bd,
@@ -301,6 +301,9 @@ def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
elif dtype in ['fp8', 'bf8']:
# rope/paged-kv is not supported
pipelines.append(FmhaFwdAppendKVPipeline('col', 't', 't', 't', 't', 'no', 'f'))
elif dtype in ['fp8fp16', 'fp8bf16']:
# TODO
None
else:
assert False
return pipelines
@@ -308,7 +311,7 @@ def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
gen = list()
api_pool = FmhaFwdAppendKVApiPool(mask_impl)
for dtype in DTYPE_MAP.keys():
for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_appendkv_tile_dict_from_dtype(dtype)
if d == None:
continue

73
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View File

@@ -112,7 +112,7 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
}}
using trait_{F_idx} = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
{F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad},
{F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad},
{F_dvpad}>;
#include <iostream>
@@ -161,7 +161,7 @@ using fmha_pipeline_problem = ck_tile::BlockFmhaSplitKVCombinePipelineProblem<
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
{F_hdim},
{F_bm0},
{F_bm0},
{F_bn1},
{F_mode},
fmha_trait>;
@@ -231,11 +231,11 @@ float fmha_fwd_splitkv_(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a
if(s.log_level_ > 0)
std::cout
<< ", " << fmha_fwd_splitkv_get_name_<fmha_fwd_splitkv_traits_>()
<< ", " << fmha_fwd_splitkv_combine_get_name_<fmha_fwd_splitkv_combine_traits_>()
<< ", " << fmha_fwd_splitkv_combine_get_name_<fmha_fwd_splitkv_combine_traits_>()
<< std::flush;
return ck_tile::launch_kernel(s,
[=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_oneshot_<fmha_fwd_splitkv_traits_>(s_, a); }},
[=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_oneshot_<fmha_fwd_splitkv_traits_>(s_, a); }},
[=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_combine_oneshot_<fmha_fwd_splitkv_combine_traits_>(s_, a); }}
);
}}
@@ -247,12 +247,22 @@ float fmha_fwd_splitkv(fmha_fwd_splitkv_traits t, fmha_fwd_splitkv_args a, const
}}
"""
FMHA_FWD_SPLITKV_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse}) && (t.do_fp8_static_quant == {F_squant}) &&
FMHA_FWD_SPLITKV_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.do_fp8_static_quant == {F_squant}) &&
((a.block_table_ptr != nullptr) == {F_pagedkv}) && ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, true, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
if (t.has_lse) {{
if constexpr (std::is_same_v<{F_dtype}, ck_tile::fp8_t>) {{
return -1;
}} else {{
using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, true, {F_squant}, {F_spad}, {F_dvpad}>;
return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
}}
}} else {{
using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, false, {F_squant}, {F_spad}, {F_dvpad}>;
return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
}}
}}
"""
@@ -421,11 +431,11 @@ class FmhaFwdSplitKVApiPool:
inners = inners + FMHA_FWD_SPLITKV_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias],
F_lse=BOOL_MAP[trait.lse], F_squant=BOOL_MAP[trait.squant], F_pagedkv=BOOL_MAP[trait.pagedkv],
F_lse=BOOL_MAP[trait.lse], F_squant=BOOL_MAP[trait.squant], F_pagedkv=BOOL_MAP[trait.pagedkv],
F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
@@ -462,7 +472,7 @@ class FmhaFwdSplitKVKernel:
FMHA_FWD_SPLITKV_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0,
@@ -482,7 +492,7 @@ class FmhaFwdSplitKVKernel:
F_spad = BOOL_MAP[self.F_pipeline.F_spad],
F_skpad = BOOL_MAP[self.F_pipeline.F_skpad],
F_dpad = BOOL_MAP[self.F_pipeline.F_dpad],
F_dvpad = BOOL_MAP[self.F_pipeline.F_dvpad],
F_dvpad = BOOL_MAP[self.F_pipeline.F_dvpad],
F_bias = BIAS_MAP[self.F_pipeline.F_bias],
F_lse = BOOL_MAP[self.F_pipeline.F_lse],
F_squant = BOOL_MAP[self.F_pipeline.F_squant],
@@ -542,7 +552,7 @@ class FmhaFwdSplitKVCombineKernel:
FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0,
F_bn1 = self.F_tile.F_bn1,
F_spad = BOOL_MAP[self.F_pipeline.F_spad],
@@ -614,27 +624,29 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
squant = 't' if dtype == 'fp8' else 'f'
pipelines = []
if dtype in ['fp16', 'bf16']:
for mask, bias, lse, pagedkv in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]):
# TODO: use async pipeline when compiler is more stable
for mask, bias, pagedkv in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"]):
# TODO: use async pipeline when compiler is more stable
if hdim == 256 or hdim in [32, 64, 128]: ### [32, 64, 96, 128]:
# if True:
pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, 't', squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, 't', squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
else:
pipelines.append(Pipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'row', 't', 'f', 't', 't', bias, 't', squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'row', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'col', 't', 'f', 't', 't', bias, 't', squant, pagedkv, mask))
pipelines.append(Pipeline('qr_async', 'col', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
if receipt == 1:
pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask)) # TODO: cover arbitraty hdim
pipelines.append(Pipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, squant, pagedkv, mask)) # TODO: cover arbitraty hdim
pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask)) # TODO: cover arbitraty hdim
pipelines.append(Pipeline('qr', 'col', 't', 'f', 't', 't', bias, 't', squant, pagedkv, mask)) # TODO: cover arbitraty hdim
elif dtype in ['fp8', 'bf8']:
# no need lse/paged-kv kernels
for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', squant, 'f', mask))
pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 't', squant, 'f', mask))
elif dtype in ['fp8fp16', 'fp8bf16']:
# TODO
None
else:
assert False
return pipelines
@@ -642,7 +654,7 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
gen = list()
api_pool = FmhaFwdSplitKVApiPool(mask_impl)
for dtype in DTYPE_MAP.keys():
for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_tile_dict_from_dtype(dtype)
if d == None:
continue
@@ -655,9 +667,6 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
# in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
continue
if pipeline.F_pagedkv == 't':
# we only use batch mode kernels to handle (paged-) kvcache problems
continue
k = Kernel(F_idx=0,
F_hdim=hdim,
F_dtype=dtype,
@@ -705,7 +714,7 @@ def get_fwd_splitkv_combine_blobs(kernel_filter : Optional[str], receipt) -> Lis
gen = list()
for dtype in DTYPE_MAP.keys():
for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype)
if d == None:
continue

14
example/ck_tile/01_fmha/fmha_bwd.cpp
View File

@@ -101,7 +101,7 @@ auto create_args(int argc, char* argv[])
}
// different threshold for different dtype
template <typename DataType>
template <typename DataTypeConfig>
auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/)
{
double rtol = 1e-2;
@@ -110,7 +110,7 @@ auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/)
}
template <>
auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_v)
auto get_elimit<FmhaBwdBf16>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_v)
{
double rtol = 1e-2;
double atol = 1e-2;
@@ -122,7 +122,7 @@ auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_
return ck_tile::make_tuple(rtol, atol);
}
template <typename DataType>
template <typename DataTypeConfig>
bool run(const ck_tile::ArgParser& arg_parser)
{
std::string data_type = arg_parser.get_str("prec");
@@ -209,7 +209,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
const auto seqstart_q_host = generate_seqstarts(mode, batch, seqlen_q);
const auto seqstart_k_host = generate_seqstarts(mode, batch, seqlen_k);
using TypeConfig = FmhaBwdTypeConfig<DataType>;
using TypeConfig = FmhaBwdTypeConfig<DataTypeConfig>;
using QDataType = typename TypeConfig::QDataType;
using KDataType = typename TypeConfig::KDataType;
@@ -933,7 +933,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
// clang-format on
auto [rtol, atol] = get_elimit<DataType>(hdim_q, hdim_v);
auto [rtol, atol] = get_elimit<DataTypeConfig>(hdim_q, hdim_v);
bool dq_cur_pass = ck_tile::check_err(dq_host_result,
dq_host_ref,
std::string("Error: QGrad Incorrect results!"),
@@ -986,11 +986,11 @@ int main(int argc, char* argv[])
const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16")
{
return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
return run<FmhaBwdFp16>(arg_parser) ? 0 : -2;
}
else if(data_type == "bf16")
{
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
return run<FmhaBwdBf16>(arg_parser) ? 0 : -2;
}
return -3;

220
example/ck_tile/01_fmha/fmha_bwd.hpp
View File

@@ -14,11 +14,19 @@
#include <utility>
#include <variant>
struct FmhaBwdFp16
{
};
struct FmhaBwdBf16
{
};
template <typename DataType>
struct FmhaBwdTypeConfig;
template <>
struct FmhaBwdTypeConfig<ck_tile::half_t>
struct FmhaBwdTypeConfig<FmhaBwdFp16>
{
using QDataType = ck_tile::half_t;
using KDataType = ck_tile::half_t;
@@ -38,7 +46,7 @@ struct FmhaBwdTypeConfig<ck_tile::half_t>
};
template <>
struct FmhaBwdTypeConfig<ck_tile::bf16_t>
struct FmhaBwdTypeConfig<FmhaBwdBf16>
{
using QDataType = ck_tile::bf16_t;
using KDataType = ck_tile::bf16_t;
@@ -150,113 +158,113 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
// create group mode kernel arguments
if constexpr(FmhaBwdDQDKDVKernel::kIsGroupMode)
{
return FmhaBwdDQDKDVKernel::MakeKargs(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.lse_ptr,
args.do_ptr,
args.d_ptr,
args.rand_val_ptr,
args.dk_ptr,
args.dv_ptr,
args.dbias_ptr,
args.dq_acc_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.seqlen_k_ptr,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_do,
args.stride_dq_acc,
args.stride_dk,
args.stride_dv,
args.stride_dbias,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_do,
args.nhead_stride_lsed,
args.nhead_stride_dq_acc,
args.nhead_stride_dk,
args.nhead_stride_dv,
args.nhead_stride_dbias,
args.split_stride_dq_acc,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.drop_seed_offset);
return FmhaBwdDQDKDVKernel::MakeKargsImpl(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.lse_ptr,
args.do_ptr,
args.d_ptr,
args.rand_val_ptr,
args.dk_ptr,
args.dv_ptr,
args.dbias_ptr,
args.dq_acc_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.seqlen_k_ptr,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_do,
args.stride_dq_acc,
args.stride_dk,
args.stride_dv,
args.stride_dbias,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_do,
args.nhead_stride_lsed,
args.nhead_stride_dq_acc,
args.nhead_stride_dk,
args.nhead_stride_dv,
args.nhead_stride_dbias,
args.split_stride_dq_acc,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.drop_seed_offset);
}
else
{ // create batch mode kernel arguments
return FmhaBwdDQDKDVKernel::MakeKargs(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.lse_ptr,
args.do_ptr,
args.d_ptr,
args.rand_val_ptr,
args.dk_ptr,
args.dv_ptr,
args.dbias_ptr,
args.dq_acc_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_do,
args.stride_dq_acc,
args.stride_dk,
args.stride_dv,
args.stride_dbias,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_do,
args.nhead_stride_lsed,
args.nhead_stride_dq_acc,
args.nhead_stride_dk,
args.nhead_stride_dv,
args.nhead_stride_dbias,
args.batch_stride_q,
args.batch_stride_k,
args.batch_stride_v,
args.batch_stride_bias,
args.batch_stride_randval,
args.batch_stride_do,
args.batch_stride_lsed,
args.batch_stride_dq_acc,
args.batch_stride_dk,
args.batch_stride_dv,
args.batch_stride_dbias,
args.split_stride_dq_acc,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.drop_seed_offset);
return FmhaBwdDQDKDVKernel::MakeKargsImpl(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.lse_ptr,
args.do_ptr,
args.d_ptr,
args.rand_val_ptr,
args.dk_ptr,
args.dv_ptr,
args.dbias_ptr,
args.dq_acc_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_do,
args.stride_dq_acc,
args.stride_dk,
args.stride_dv,
args.stride_dbias,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_do,
args.nhead_stride_lsed,
args.nhead_stride_dq_acc,
args.nhead_stride_dk,
args.nhead_stride_dv,
args.nhead_stride_dbias,
args.batch_stride_q,
args.batch_stride_k,
args.batch_stride_v,
args.batch_stride_bias,
args.batch_stride_randval,
args.batch_stride_do,
args.batch_stride_lsed,
args.batch_stride_dq_acc,
args.batch_stride_dk,
args.batch_stride_dv,
args.batch_stride_dbias,
args.split_stride_dq_acc,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.drop_seed_offset);
}
}();

218
example/ck_tile/01_fmha/fmha_fwd.cpp
View File

@@ -3,6 +3,7 @@
#include "fmha_fwd.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/ref/naive_attention.hpp"
#include "mask.hpp"
#include "rotary.hpp"
#include "utils.hpp"
@@ -41,7 +42,7 @@ std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("v", "1", "weather do CPU validation or not")
arg_parser.insert("v", "1", "0:no validation, 2:cpu validation, 2:gpu validation(experimental)")
.insert("mode", "0", "kernel mode. 0:batch, 1:group")
.insert("b", "2", "batch size")
.insert("h", "8", "num of head, for q")
@@ -62,7 +63,7 @@ auto create_args(int argc, char* argv[])
"-1 to choose s_knew in [1, s] randomly.")
.insert("s_kpad",
"-1",
"seqlen_k stride between 2 tokens, currently used in group-mode only\n"
"seqlen_k stride between 2 batches, currently used in group-mode only\n"
"for kv-cache case, each batch [1,s,h,d]/[1,h,s,d] can have a stride\n"
"along seqlen, instead of packed. same as xformer kv_padding")
.insert("d", "128", "head dim for q, k")
@@ -142,7 +143,7 @@ auto create_args(int argc, char* argv[])
}
// different threshold for different dtype
template <typename DataType>
template <typename DataTypeConfig>
auto get_elimit(std::string /*init_method*/)
{
double rtol = 1e-3;
@@ -151,7 +152,7 @@ auto get_elimit(std::string /*init_method*/)
}
template <>
auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/)
auto get_elimit<FmhaFwdBf16>(std::string /*init_method*/)
{
double rtol = 1e-2;
double atol = 1e-2;
@@ -159,7 +160,7 @@ auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/)
}
template <>
auto get_elimit<ck_tile::fp8_t>(std::string init_method)
auto get_elimit<FmhaFwdFp8>(std::string init_method)
{
if(init_method == "ui" || init_method == "ni")
{
@@ -261,7 +262,7 @@ int override_num_splits_if_necessary(
return num_splits;
}
template <typename DataType>
template <typename DataTypeConfig>
bool run(const ck_tile::ArgParser& arg_parser)
{
std::string data_type = arg_parser.get_str("prec");
@@ -294,7 +295,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
#if !CK_TILE_FMHA_FWD_APPENDKV_API
if(seqlen_knew != 0)
{
std::cerr << "kvcache is not supported. ignoring the 's_knew' option" << std::endl;
std::cerr << "fmha_fwd_appendkv() is not enabled. ignoring the 's_knew' option"
<< std::endl;
seqlen_knew = 0;
}
#endif
@@ -304,8 +306,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
ck_tile::index_t rotary_dim = arg_parser.get_int("rotary_dim");
if constexpr(!(std::is_same_v<DataType, ck_tile::fp16_t> ||
std::is_same_v<DataType, ck_tile::bf16_t>))
if constexpr(!(std::is_same_v<DataTypeConfig, FmhaFwdFp16> ||
std::is_same_v<DataTypeConfig, FmhaFwdBf16>))
{
if(0 < rotary_dim)
{
@@ -321,6 +323,13 @@ bool run(const ck_tile::ArgParser& arg_parser)
rotary_dim = 0;
}
#endif
// to use fmha_fwd_appendkv(), make sure it's in batch mode
const bool need_append_kvcache = (0 < seqlen_knew || 0 < rotary_dim);
if(need_append_kvcache && mode == mode_enum::group)
{
std::cerr << "fmha_fwd_appendkv() will be invoked. ignoring the 'mode' option" << std::endl;
mode = mode_enum::batch;
}
if(!(rotary_dim <= hdim_q))
{
std::cerr << "rotary_dim should be less than or equal to head dim for q" << std::endl;
@@ -356,22 +365,26 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< std::endl;
use_cache_batch_idx = false;
}
#else
if(use_cache_batch_idx)
{
if(0 < page_block_size)
{
std::cerr << "paged-kvcache does not support cache_batch_idx. ignoring the "
"'cache_batch_idx' option"
<< std::endl;
use_cache_batch_idx = false;
}
else if(mode == mode_enum::group)
{
std::cerr << "group mode will not use cache_batch_idx. ignoring the "
"'cache_batch_idx' option"
<< std::endl;
use_cache_batch_idx = false;
}
}
#endif
if(0 < page_block_size && use_cache_batch_idx)
{
std::cerr << "paged-kvcache does not support cache_batch_idx. ignoring the "
"'cache_batch_idx' option"
<< std::endl;
use_cache_batch_idx = false;
}
// the input tensor layout for kvcache is same as batch mode
const bool need_append_kvcache = (0 < seqlen_knew || 0 < rotary_dim);
const bool use_kvcache = (need_append_kvcache || use_cache_batch_idx || 0 < page_block_size);
if(use_kvcache && mode != mode_enum::batch)
{
std::cerr << "kvcache enabled. ignoring the 'mode' option" << std::endl;
mode = mode_enum::batch;
}
auto [seqlen_qs, seqlen_ks, seqlen_kpads] =
decode_seqlen(mode,
@@ -380,7 +393,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
arg_parser.get_str("s_k"),
arg_parser.get_str("s_kpad"),
/*seqlen_k_min=*/0 < seqlen_knew ? seqlen_knew : 0,
use_kvcache);
need_append_kvcache);
// compute kvcache seqlen_k (before appending knew/vnew)
auto cache_seqlen_ks = seqlen_ks;
std::transform(cache_seqlen_ks.begin(),
@@ -416,25 +429,6 @@ bool run(const ck_tile::ArgParser& arg_parser)
return atoi(squant_str.c_str()) != 0 ? true : false;
}();
float range_q = arg_parser.get_float("range_q");
float range_k = arg_parser.get_float("range_k");
float range_v = arg_parser.get_float("range_v");
float range_p = arg_parser.get_float("range_p");
float range_o = arg_parser.get_float("range_o");
float dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<DataType>::max());
float scale_p = 1.f;
float scale_o = 1.f;
if(squant)
{
scale_s = scale_s * (range_q / dtype_max) * (range_k / dtype_max);
scale_p = dtype_max / range_p;
// scale_p = [max(fp8_t)/range_o] * [range_p/max(fp8_t)] * [range_v/max(fp8_t)]
scale_o = range_p * range_v / range_o / dtype_max;
}
std::string vlayout = arg_parser.get_str("vlayout");
bool lse = arg_parser.get_bool("lse");
@@ -454,7 +448,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
bool s_randval = false;
if(p_drop > 0.0f && do_validation)
if(p_drop > 0.0f && do_validation != 0)
{
s_randval = true;
}
@@ -487,7 +481,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
const auto seqstart_k_host = to_seqstarts(seqlen_ks);
const auto seqstart_k_with_padding_host = to_seqstarts(seqlen_kpads);
using TypeConfig = FmhaFwdTypeConfig<DataType>;
using TypeConfig = FmhaFwdTypeConfig<DataTypeConfig>;
using QDataType = typename TypeConfig::QDataType;
using KDataType = typename TypeConfig::KDataType;
@@ -501,6 +495,28 @@ bool run(const ck_tile::ArgParser& arg_parser)
using OaccDataType = typename TypeConfig::OaccDataType;
using ODataType = typename TypeConfig::ODataType;
float range_q = arg_parser.get_float("range_q");
float range_k = arg_parser.get_float("range_k");
float range_v = arg_parser.get_float("range_v");
float range_p = arg_parser.get_float("range_p");
float range_o = arg_parser.get_float("range_o");
float q_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<QDataType>::max());
float k_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<KDataType>::max());
float v_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<VDataType>::max());
float p_dtype_max = v_dtype_max; // assume p and v is the same type
float o_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<ODataType>::max());
float scale_p = 1.f;
float scale_o = 1.f;
if(squant)
{
scale_s = scale_s * (range_q / q_dtype_max) * (range_k / k_dtype_max);
scale_p = p_dtype_max / range_p;
scale_o = (o_dtype_max / range_o) * (range_p / p_dtype_max) * (range_v / v_dtype_max);
}
// accumulation numbers for performance evaluation
std::size_t flop = 0, num_byte = 0;
auto max_seqlen_q =
@@ -697,14 +713,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
else if(init_method == "ufq" || init_method == "uf:q" ||
init_method == "3") // suitable for fp8 quantization
{
ck_tile::FillUniformDistribution<QDataType>{-dtype_max, dtype_max, seed}(q_host);
ck_tile::FillUniformDistribution<KDataType>{-dtype_max, dtype_max, seed}(k_host);
ck_tile::FillUniformDistribution<KDataType>{-dtype_max, dtype_max, seed}(knew_host);
ck_tile::FillUniformDistribution<VDataType>{-dtype_max, dtype_max, seed}(v_host);
ck_tile::FillUniformDistribution<VDataType>{-dtype_max, dtype_max, seed}(vnew_host);
ck_tile::FillUniformDistribution<QDataType>{-q_dtype_max, q_dtype_max, seed}(q_host);
ck_tile::FillUniformDistribution<KDataType>{-k_dtype_max, k_dtype_max, seed}(k_host);
ck_tile::FillUniformDistribution<KDataType>{-k_dtype_max, k_dtype_max, seed}(knew_host);
ck_tile::FillUniformDistribution<VDataType>{-v_dtype_max, v_dtype_max, seed}(v_host);
ck_tile::FillUniformDistribution<VDataType>{-v_dtype_max, v_dtype_max, seed}(vnew_host);
// bias_fp8 = qscale_bias * bias_fp32
float qscale_bias = (dtype_max / range_q) * (dtype_max / range_k);
float qscale_bias = (q_dtype_max / range_q) * (k_dtype_max / range_k);
// Assume bias is in [-1.f, 1.f] in original fp32
ck_tile::FillUniformDistribution<BiasDataType>{-qscale_bias, qscale_bias, seed}(bias_host);
}
@@ -741,8 +757,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
ck_tile::DeviceMem seqlen_k_buf(
use_kvcache || 0 <= seqlen_kpads[0] ? seqlen_ks.size() * sizeof(int32_t) : 0);
ck_tile::DeviceMem seqlen_k_buf((mode == mode_enum::batch && use_kvcache) ||
0 <= seqlen_kpads[0]
? seqlen_ks.size() * sizeof(int32_t)
: 0);
ck_tile::DeviceMem cache_seqlen_k_buf(
need_append_kvcache ? cache_seqlen_ks.size() * sizeof(int32_t) : 0);
ck_tile::DeviceMem rotary_cos_buf(rotary_cos_host.get_element_space_size_in_bytes());
@@ -763,7 +781,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
seqstart_q.ToDevice(seqstart_q_host.data());
seqstart_k.ToDevice(seqlen_kpads[0] < 0 ? seqstart_k_host.data()
: seqstart_k_with_padding_host.data());
seqlen_k_buf.ToDevice(use_kvcache || 0 <= seqlen_kpads[0] ? seqlen_ks.data() : nullptr);
seqlen_k_buf.ToDevice((mode == mode_enum::batch && use_kvcache) || 0 <= seqlen_kpads[0]
? seqlen_ks.data()
: nullptr);
cache_seqlen_k_buf.ToDevice(need_append_kvcache ? cache_seqlen_ks.data() : nullptr);
rotary_cos_buf.ToDevice(rotary_cos_host.data());
rotary_sin_buf.ToDevice(rotary_sin_host.data());
@@ -976,8 +996,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
(mode == mode_enum::group ? seqstart_q.GetDeviceBuffer() : nullptr);
args.seqstart_k_ptr =
(mode == mode_enum::group ? seqstart_k.GetDeviceBuffer() : nullptr);
args.seqlen_k_ptr =
(use_kvcache || 0 <= k_paddings_[0] ? seqlen_k_buf.GetDeviceBuffer() : nullptr);
args.seqlen_k_ptr = ((mode == mode_enum::batch && use_kvcache) || 0 <= k_paddings_[0]
? seqlen_k_buf.GetDeviceBuffer()
: nullptr);
args.seqlen_k = shape_seqlen_k; // unused in group mode (or kvcache enabled)
args.max_seqlen_q = max_seqlen_q;
@@ -1029,6 +1050,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
(0 < page_block_size ? block_table_buf.GetDeviceBuffer() : nullptr);
args.batch_stride_block_table = batch_stride_block_table;
args.page_block_size = page_block_size;
args.is_gappy = false; // use 'false' for flash-attention integration
args.cache_batch_idx =
(use_cache_batch_idx ? cache_batch_idx_buf.GetDeviceBuffer() : nullptr);
@@ -1100,25 +1122,75 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< std::setprecision(2) << tflops << " TFlops, " << std::setprecision(2) << gb_per_sec
<< " GB/s" << std::flush;
if(!do_validation)
if(do_validation == 0)
{
std::cout << std::flush << std::endl;
return true;
}
if(do_validation == 2)
{
// NOTE: use gpu to do validation
ck_tile::naive_attention_fwd_traits naive_t;
naive_t.q_type = data_type;
naive_t.k_type = data_type;
naive_t.v_type = data_type;
naive_t.o_type = data_type;
naive_t.q_layout = i_perm == 1 ? "bhsd" : "bshd";
naive_t.k_layout = i_perm == 1 ? "bhsd" : "bshd";
naive_t.v_layout = i_perm == 1 ? "bhsd" : "bshd";
naive_t.o_layout = o_perm == 1 ? "bhsd" : "bshd";
naive_t.variation = 0; // TODO?
ck_tile::DeviceMem o_naive_buf(o_host.get_element_space_size_in_bytes());
ck_tile::naive_attention_fwd_args naive_a;
naive_a.q_ptr = q_buf.GetDeviceBuffer();
naive_a.k_ptr = k_buf.GetDeviceBuffer();
naive_a.v_ptr = v_buf.GetDeviceBuffer();
naive_a.o_ptr = o_naive_buf.GetDeviceBuffer();
naive_a.scale_s = scale_s;
naive_a.context_len_ptr = nullptr; // used when seqlen kv come from a pointer
naive_a.page_table_ptr =
nullptr; // [batch, num_blocks] seqlen_kv is in different block(paged attn)
naive_a.hdim = hdim_q;
naive_a.hdim_v = hdim_v; // could be cross-attn, where V and Q/K hdim are different
naive_a.batch_q = batch;
naive_a.batch_kv = batch;
naive_a.batch_ratio_kv = 1; // batch_q / batch_kv
naive_a.seqlen_q = seqlen_qs[0];
naive_a.seqlen_kv = seqlen_ks[0]; // if context_len_ptr is not nullptr, ignore this field
naive_a.nhead_q = nhead;
naive_a.nhead_kv = nhead_k;
naive_a.nhead_ratio_kv = naive_a.nhead_q / naive_a.nhead_kv; // nhead_q / nhead_kv
naive_a.page_size = 0; // if paged, the seqlen-kv for each block
ck_tile::stream_config naive_s{};
naive_attention_fwd(naive_t, naive_a, naive_s);
auto o_naive_ref = o_naive_buf.ToHost<ODataType>();
o_buf.FromDevice(o_host.data()); // TODO: ugly
auto [rtol_, atol_] = get_elimit<DataTypeConfig>(init_method);
bool pass_ = ck_tile::check_err(
o_host, o_naive_ref, std::string("OUT Error: Incorrect results!"), rtol_, atol_);
std::cout << ", valid:" << (pass_ ? "y" : "n") << std::flush << std::endl;
return pass_;
}
o_buf.FromDevice(o_host.data());
lse_buf.FromDevice(lse_host.data());
randval_buf.FromDevice(randval_host.data());
auto p_compute_element_func = [&]() {
if constexpr(std::is_same_v<DataType, ck_tile::fp8_t>)
if constexpr(std::is_same_v<DataTypeConfig, ck_tile::fp8_t>)
return ck_tile::scales{scale_p};
else
return ck_tile::identity{};
}();
auto oacc_element_func = [&]() {
if constexpr(std::is_same_v<DataType, ck_tile::fp8_t>)
if constexpr(std::is_same_v<DataTypeConfig, ck_tile::fp8_t>)
return ck_tile::composes(ck_tile::saturates<ck_tile::fp8_t>{},
ck_tile::scales{scale_o});
else
@@ -1168,7 +1240,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
decltype(q_host_ref) q_host_ref_ro(q_host_ref.get_lengths());
auto [rotary_cos_slice, rotary_sin_slice] =
auto [rotary_cos_slice, rotary_sin_slice] =
slice_rotary_cos_sin(rotary_cos_host, rotary_sin_host, cache_seqlen_ks[wb], real_seqlen_q);
ck_tile::reference_batched_rotary_position_embedding(
@@ -1184,13 +1256,13 @@ bool run(const ck_tile::ArgParser& arg_parser)
k_host_ref.ForEach([&](auto& self, auto i) {
self(i) = k_host(block_table_host(wb, i[1] / page_block_size), i[0] / nr, i[1] % page_block_size, i[2]);
});
} else {
} else {
k_host_ref.ForEach([&](auto& self, auto i) {
self(i) = k_host(block_table_host(wb, i[1] / page_block_size), i[1] % page_block_size, i[0] / nr, i[2]);
});
}
} else
#endif
#endif
{
if(i_perm) k_host_ref.ForEach([&](auto& self, auto i) { self(i) = k_host(cache_b_idx, i[0] / nr, i[1] + key_offset, i[2]); });
else k_host_ref.ForEach([&](auto& self, auto i) { self(i) = k_host(cache_b_idx, i[1] + key_offset, i[0] / nr, i[2]); });
@@ -1211,7 +1283,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
knew_host_ref_ro.emplace(knew_host_ref.get_lengths());
auto [rotary_cos_slice, rotary_sin_slice] =
auto [rotary_cos_slice, rotary_sin_slice] =
slice_rotary_cos_sin(rotary_cos_host, rotary_sin_host, cache_seqlen_ks[wb], seqlen_knew);
ck_tile::reference_batched_rotary_position_embedding(
@@ -1233,19 +1305,19 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(0 < page_block_size) {
if(is_v_rowmajor) {
if(i_perm) {
v_host_ref.ForEach([&](auto& self, auto i) {
self(i) = v_host(block_table_host(wb, i[2] / page_block_size), i[0] / nr, i[2] % page_block_size, i[1]);
v_host_ref.ForEach([&](auto& self, auto i) {
self(i) = v_host(block_table_host(wb, i[2] / page_block_size), i[0] / nr, i[2] % page_block_size, i[1]);
});
} else {
v_host_ref.ForEach([&](auto& self, auto i) {
v_host_ref.ForEach([&](auto& self, auto i) {
self(i) = v_host(block_table_host(wb, i[2] / page_block_size), i[2] % page_block_size, i[0] / nr, i[1]);
});
}
}
else
else
{
if(i_perm) {
v_host_ref.ForEach([&](auto& self, auto i) {
if(i_perm) {
v_host_ref.ForEach([&](auto& self, auto i) {
self(i) = v_host(block_table_host(wb, i[2] / page_block_size), i[0] / nr, i[1], i[2] % page_block_size);
});
} else {
@@ -1440,7 +1512,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
else o_host_result.ForEach([&](auto& self, auto idx) { self(idx) = o_host(b_idx, idx[1] + query_offset, idx[0], idx[2]); });
// clang-format on
auto [rtol, atol] = get_elimit<DataType>(init_method);
auto [rtol, atol] = get_elimit<DataTypeConfig>(init_method);
bool cur_pass = ck_tile::check_err(
o_host_result, o_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
pass &= cur_pass;
@@ -1497,15 +1569,15 @@ int main(int argc, char* argv[])
const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16")
{
return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
return run<FmhaFwdFp16>(arg_parser) ? 0 : -2;
}
else if(data_type == "bf16")
{
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
return run<FmhaFwdBf16>(arg_parser) ? 0 : -2;
}
else if(data_type == "fp8")
{
return run<ck_tile::fp8_t>(arg_parser) ? 0 : -2;
return run<FmhaFwdFp8>(arg_parser) ? 0 : -2;
}
return -3;

210
example/ck_tile/01_fmha/fmha_fwd.hpp
View File

@@ -16,11 +16,35 @@
#include <utility>
#include <variant>
struct FmhaFwdFp16
{
};
struct FmhaFwdBf16
{
};
struct FmhaFwdFp8
{
};
struct FmhaFwdBf8
{
};
struct FmhaFwdFp8Fp16
{
};
struct FmhaFwdFp8Bf16
{
};
template <typename DataType>
struct FmhaFwdTypeConfig;
template <>
struct FmhaFwdTypeConfig<ck_tile::half_t>
struct FmhaFwdTypeConfig<FmhaFwdFp16>
{
using QDataType = ck_tile::half_t;
using KDataType = ck_tile::half_t;
@@ -36,7 +60,7 @@ struct FmhaFwdTypeConfig<ck_tile::half_t>
};
template <>
struct FmhaFwdTypeConfig<ck_tile::bf16_t>
struct FmhaFwdTypeConfig<FmhaFwdBf16>
{
using QDataType = ck_tile::bf16_t;
using KDataType = ck_tile::bf16_t;
@@ -52,7 +76,7 @@ struct FmhaFwdTypeConfig<ck_tile::bf16_t>
};
template <>
struct FmhaFwdTypeConfig<ck_tile::fp8_t>
struct FmhaFwdTypeConfig<FmhaFwdFp8>
{
using QDataType = ck_tile::fp8_t;
using KDataType = ck_tile::fp8_t;
@@ -68,7 +92,7 @@ struct FmhaFwdTypeConfig<ck_tile::fp8_t>
};
template <>
struct FmhaFwdTypeConfig<ck_tile::bf8_t>
struct FmhaFwdTypeConfig<FmhaFwdBf8>
{
using QDataType = ck_tile::bf8_t;
using KDataType = ck_tile::bf8_t;
@@ -165,6 +189,8 @@ struct fmha_fwd_splitkv_args
void* block_table_ptr;
ck_tile::index_t batch_stride_block_table; // only used if 'block_table_ptr' is not nullptr
ck_tile::index_t page_block_size; // only used if 'block_table_ptr' is not nullptr
bool is_gappy; // differentiate seqstart_k_ptr usage. only used if 'block_table_ptr' is not
// nullptr.
const void* cache_batch_idx;
@@ -173,9 +199,21 @@ struct fmha_fwd_splitkv_args
// seqlen_k = kargs.seqlen_k
// group mode: seqlen_q = kargs.seqstart_q_ptr[b + 1] - kargs.seqstart_q_ptr[b]
// seqlen_k = kargs.seqstart_k_ptr[b + 1] - kargs.seqstart_k_ptr[b]
// kvcache mode (use same kernel as batch mode):
// or kargs.seqlen_k_ptr[b]
//
// batch mode (kvcache):
// seqlen_q = kargs.seqlen_q
// seqlen_k = kargs.seqlen_k_ptr[b]
// group mode (kvcache):
// seqlen_q = kargs.seqstart_q_ptr[b + 1] - kargs.seqstart_q_ptr[b]
//
// when is_gappy=true:
// seqlen_k = kargs.seqlen_k_ptr[b]
// seqstart_k_ptr[b] now store local offset of each batch
//
// when is_gappy=false:
// seqlen_k = kargs.seqstart_k_ptr[b + 1] - kargs.seqstart_k_ptr[b]
// or kargs.seqlen_k_ptr[b]
const void* seqstart_q_ptr;
const void* seqstart_k_ptr;
const void* seqlen_k_ptr;
@@ -251,7 +289,7 @@ struct fmha_fwd_appendkv_args
ck_tile::index_t batch_stride_block_table; // only used if 'block_table_ptr' is not nullptr
ck_tile::index_t page_block_size; // only used if 'block_table_ptr' is not nullptr
const void* cache_batch_idx;
const void* cache_batch_idx; // only used if block_table_ptr is nullptr -> batch mode (kvcache)
ck_tile::index_t stride_q;
ck_tile::index_t stride_k;
@@ -278,87 +316,87 @@ auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
// create group mode kernel arguments
if constexpr(FmhaKernel::kIsGroupMode)
{
return FmhaKernel::MakeKargs(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.rand_val_ptr,
args.lse_ptr,
args.o_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.seqlen_k_ptr,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale_s,
args.scale_p,
args.scale_o,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_o,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_lse,
args.nhead_stride_o,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
return FmhaKernel::MakeKargsImpl(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.rand_val_ptr,
args.lse_ptr,
args.o_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.seqlen_k_ptr,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale_s,
args.scale_p,
args.scale_o,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_o,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_lse,
args.nhead_stride_o,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
}
else
{ // create batch mode kernel arguments
return FmhaKernel::MakeKargs(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.rand_val_ptr,
args.lse_ptr,
args.o_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale_s,
args.scale_p,
args.scale_o,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_o,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_lse,
args.nhead_stride_o,
args.batch_stride_q,
args.batch_stride_k,
args.batch_stride_v,
args.batch_stride_bias,
args.batch_stride_randval,
args.batch_stride_lse,
args.batch_stride_o,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
return FmhaKernel::MakeKargsImpl(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.rand_val_ptr,
args.lse_ptr,
args.o_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.scale_s,
args.scale_p,
args.scale_o,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.stride_o,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.nhead_stride_lse,
args.nhead_stride_o,
args.batch_stride_q,
args.batch_stride_k,
args.batch_stride_v,
args.batch_stride_bias,
args.batch_stride_randval,
args.batch_stride_lse,
args.batch_stride_o,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
}
}();
@@ -389,6 +427,10 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
args.nhead_q,
args.nhead_q / args.nhead_k,
args.num_splits,
args.block_table_ptr,
args.batch_stride_block_table,
args.page_block_size,
args.is_gappy,
args.scale_s,
args.scale_p,
args.stride_q,

4
example/ck_tile/01_fmha/utils.hpp
View File

@@ -145,7 +145,7 @@ decode_seqlen(mode_enum mode,
std::string k_val,
std::string k_pad_val,
ck_tile::index_t seqlen_k_min = 0,
bool use_kvcache = false,
bool need_append_kvcache = false,
std::optional<unsigned> seed = std::nullopt)
{
#define _S2I_(str_) static_cast<ck_tile::index_t>(std::atoi((str_).c_str()))
@@ -159,7 +159,7 @@ decode_seqlen(mode_enum mode,
const ck_tile::index_t seqlen_k_max = (k < 0 ? q : k);
std::vector<ck_tile::index_t> seqlen_ks(batch, seqlen_k_max);
if(1 < batch && use_kvcache)
if(1 < batch && need_append_kvcache)
{
// to keep the original s_k value, we always use seqlen_k_max in first batch
randints(std::next(seqlen_ks.begin()),

2
example/ck_tile/03_gemm/CMakeLists.txt
View File

@@ -1,2 +1,2 @@
add_executable(tile_example_gemm_basic EXCLUDE_FROM_ALL gemm_basic.cpp)
add_executable(tile_example_gemm_mem_pipeline EXCLUDE_FROM_ALL gemm_mem_pipeline.cpp)
add_executable(tile_example_universal_gemm EXCLUDE_FROM_ALL universal_gemm.cpp)

5
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -92,6 +92,11 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"

23
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -31,15 +31,13 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
float ave_time = gemm_calc<ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::string op_name{"Gemm{MemBoundPipeline}"};
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte =
sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Run " << op_name << "kernel with M =" << M << " N =" << N << " K =" << K
std::cout << "Run Gemm kernel with M =" << M << " N =" << N << " K =" << K
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
<< " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
@@ -114,7 +112,6 @@ int run_gemm_example_with_layouts(int argc,
f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
// TODO: add different init types
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
@@ -202,14 +199,16 @@ int run_gemm_example(int argc, char* argv[])
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
}
// TODO: Fixme: with latest changes to GemmPipelineAGmemBGmemCRegV1DefaultPolicy below do not
// work.
// else if(a_layout == "C" && b_layout == "C")
// {
// return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
// }
// else if(a_layout == "C" && b_layout == "R")
// {
// return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
// }
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");

58
example/ck_tile/03_gemm/gemm_mem_pipeline.cpp → example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -14,12 +14,34 @@
#include "ck_tile/host.hpp"
#include "gemm_basic.hpp"
#define CK_TILE_PIPELINE_COMPUTE 1
#define CK_TILE_PIPELINE_MEMORY 2
#ifndef CK_TILE_PIPELINE_DEFAULT
#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE
#endif
template <typename ALayout, typename BLayout, typename CLayout>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
{
// ToDo: This will be modified by the codegen code later.
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
// Memory friendly for Interwave scheduler
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t N_Tile = 32;
constexpr ck_tile::index_t K_Tile = 64;
constexpr ck_tile::index_t M_Warp = 4;
constexpr ck_tile::index_t N_Warp = 1;
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;
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
// Compute friendly for Intrawave scheduler
constexpr ck_tile::index_t M_Tile = 256;
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
@@ -28,12 +50,12 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
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;
constexpr ck_tile::index_t K_Warp_Tile = 16;
#endif
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = true;
constexpr bool kPadN = true;
constexpr bool kPadK = true;
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr int kBlockPerCu = 1;
@@ -49,8 +71,11 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<
#endif
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(args.K);
@@ -63,13 +88,21 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV3<
#endif
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
ck_tile::GemmPipelineScheduler::Interwave,
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
ck_tile::GemmPipelineScheduler::Intrawave,
#endif
has_hot_loop_v,
tail_number_v>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
@@ -86,6 +119,11 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
@@ -174,8 +212,8 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
{
std::ostringstream err;
err << "When there's no hot loop, this tail number \"" << tail_num
<< "\" is not supported! " << __FILE__ << ":" << __LINE__
<< ", in function: " << __func__;
<< "\" is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
<< "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
}

4
example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp
View File

@@ -40,7 +40,7 @@ float matrix_core_swizzle(matrix_core_swizzle_traits t,
else if(t.permute.compare("0,1,3,4,2,5") == 0)
{
constexpr matrix_core_permute_style pstyle =
matrix_core_permute_style::permute_b_nr_kr_kw_nw_kv;
matrix_core_permute_style::b_nr_kr_kw_nw_kv;
using Kernel =
matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
@@ -83,7 +83,7 @@ float matrix_core_swizzle(matrix_core_swizzle_traits t,
else if(t.permute.compare("0,1,3,4,2,5") == 0)
{
constexpr matrix_core_permute_style pstyle =
matrix_core_permute_style::permute_b_nr_kr_kw_nw_kv;
matrix_core_permute_style::b_nr_kr_kw_nw_kv;
using Kernel =
matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;

12
example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp
View File

@@ -42,8 +42,8 @@ enum class matrix_core_permute_style
{
permute_b_n0_k0_n1_k1_n2_k2 = 0, // 0,1,4,2,5,3,6
permute_b_n0_n1_k0_k1_n2_k2 = 1, // 0,1,2,4,5,3,6
permute_b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5
permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv,
b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5
b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv,
};
// assume this is B matrix, originally we have batch*n*k
@@ -203,7 +203,7 @@ struct matrix_core_swizzle_kernel
else
{
// clang-format off
// permute_b_nr_kr_kw_nw_kv or permute_b_nr_kr_waveflatten
// b_nr_kr_kw_nw_kv or b_nr_kr_waveflatten
constexpr index_t Kv = Alignment;
constexpr index_t Nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
constexpr index_t Kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
@@ -332,7 +332,7 @@ struct matrix_core_swizzle_kernel
make_tuple(sequence<0>{}, sequence<1>{}));
return tmp_1;
#else
// permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv,
// b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv,
constexpr index_t kv = Alignment;
constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
@@ -376,13 +376,13 @@ struct matrix_core_swizzle_kernel
else
{
#if MERGE_2D_013425
// permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv
// b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv
return make_tile_window(dst_view,
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
{i_n * NPerBlock, i_k * KPerBlock},
get_dst_dist());
#else
// permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv
// b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv
constexpr index_t kv = Alignment;
constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;

2
example/ck_tile/06_permute/permute.cpp
View File

@@ -264,7 +264,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
if(arg_parser.get_str("perm") == std::string("0,1,3,4,2,5"))
{
// permute_b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5
// b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5
matrix_core_swizzle_traits t;
t.data_type = data_type;
t.permute = arg_parser.get_str("perm");

4
example/ck_tile/12_smoothquant/CMakeLists.txt
View File

@@ -18,7 +18,7 @@ function (add_smoothquant_example TARGET_NAME MAIN_SRC)
target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
endfunction(add_smoothquant_example TARGET_NAME MAIN_SRC)
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})
add_smoothquant_example(tile_example_smoothquant example_smoothquant.cpp)
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})

44
example/ck_tile/12_smoothquant/example_smoothquant.cpp
View File

@@ -35,7 +35,8 @@ auto create_args(int argc, char* argv[])
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n")
.insert("x_stride", "-1", "input stride per row, if -1 then equal to n")
.insert("y_stride", "-1", "output stride per row, if -1 then equal to n")
.insert("e", "1e-5", "epsilon")
.insert("v", "1", "cpu validation or not")
.insert("prec", "fp16", "precision")
@@ -49,11 +50,14 @@ auto create_args(int argc, char* argv[])
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = n;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(x_stride < 0)
x_stride = n;
ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
if(y_stride < 0)
y_stride = n;
std::string data_type = arg_parser.get_str("prec");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
@@ -68,14 +72,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
using ComputeDataType = float;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {y_stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
@@ -116,7 +120,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.GetDeviceBuffer(),
m,
n,
stride};
x_stride,
y_stride};
auto kargs = Kernel::MakeKargs(args);
@@ -133,7 +138,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation)
{
using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1});
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {y_stride, 1});
// smooth outlier
{
auto f = [&](auto n_) {
@@ -183,7 +188,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == n)
if(y_stride == n)
{
pass = ck_tile::check_err(qy_host_dev,
qy_host_ref,
@@ -195,10 +200,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
for(int i_r = 0; i_r < m; i_r++)
{
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
qy_host_dev.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
qy_host_ref.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * y_stride,
qy_host_dev.begin() + i_r * y_stride +
n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * y_stride,
qy_host_ref.begin() + i_r * y_stride +
n);
pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) +
@@ -210,8 +217,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride
<< ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
<< " m:" << m << ", n:" << n << ", x_stride:" << x_stride
<< ", y_stride:" << y_stride << ", valid:" << (pass ? "y" : "n") << std::flush
<< std::endl;
}
return pass;

44
example/ck_tile/12_smoothquant/smoothquant.cpp
View File

@@ -33,7 +33,8 @@ auto create_args(int argc, char* argv[])
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n")
.insert("x_stride", "-1", "input stride per row, if -1 then equal to n")
.insert("y_stride", "-1", "output stride per row, if -1 then equal to n")
.insert("v", "1", "cpu validation or not")
.insert("kname", "1", "print kernel name or not")
.insert("prec", "fp16", "precision")
@@ -47,18 +48,21 @@ auto create_args(int argc, char* argv[])
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = n;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(x_stride < 0)
x_stride = n;
ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
if(y_stride < 0)
y_stride = n;
std::string data_type = arg_parser.get_str("prec");
int kname = arg_parser.get_int("kname");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
assert(stride >= n);
assert(x_stride >= n);
using TypeConfig = SmoothquantTypeConfig<DataType>;
@@ -69,14 +73,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {y_stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
@@ -90,7 +94,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
xscale_buf.ToDevice(xscale_host.data());
std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride << std::flush;
<< " m:" << m << ", n:" << n << ", x_stride:" << x_stride << ", y_stride:" << y_stride
<< std::flush;
smoothquant_traits traits{data_type};
@@ -100,7 +105,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.GetDeviceBuffer(),
m,
n,
stride};
x_stride,
y_stride};
float ave_time = smoothquant(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
@@ -116,7 +122,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation)
{
using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1});
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {y_stride, 1});
// smooth outlier
{
auto f = [&](auto n_) {
@@ -166,7 +172,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == n)
if(y_stride == n)
{
pass = ck_tile::check_err(qy_host_dev,
qy_host_ref,
@@ -178,10 +184,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
for(int i_r = 0; i_r < m; i_r++)
{
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
qy_host_dev.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
qy_host_ref.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * y_stride,
qy_host_dev.begin() + i_r * y_stride +
n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * y_stride,
qy_host_ref.begin() + i_r * y_stride +
n);
pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) +

2
example/ck_tile/13_moe_sorting/moe_sorting_api.hpp
View File

@@ -5,7 +5,7 @@
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/ops/moe_sorting.hpp"
#include "ck_tile/ops/fused_moe.hpp"
struct moe_sorting_trait
{

25
example/ck_tile/14_moe_smoothquant/CMakeLists.txt Normal file
View File

@@ -0,0 +1,25 @@
function (add_moe_smoothquant_example TARGET_NAME MAIN_SRC)
message("adding ${TARGET_NAME}")
# not using add_example_executable() to add target, since we don't want this to have
# to be included in "make all/install/check"
add_executable(${TARGET_NAME} EXCLUDE_FROM_ALL ${MAIN_SRC})
target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
foreach(source IN LISTS ARGN)
list(APPEND INSTANCE_SRCS ${source})
endforeach()
target_sources(${TARGET_NAME} PRIVATE ${INSTANCE_SRCS})
set(COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
# list(APPEND COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
endfunction(add_moe_smoothquant_example TARGET_NAME MAIN_SRC)
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_moe_smoothquant_example(tile_example_moe_smoothquant moe_smoothquant.cpp ${INSTANCE_SRCS})

15
example/ck_tile/14_moe_smoothquant/README.md Normal file
View File

@@ -0,0 +1,15 @@
# moe-smoothquant
This folder contains example for moe-smoothquant using ck_tile tile-programming implementation.
![](misc/moe-sm.png)
Unlike standard smoothquant op, the input scale is from different expert `[expert, hidden]`, we need reuse the `topk-id` from previous `topk-softmax` and select the corresponding `expert` from current topk, and expand the output/per-token-scale by `topk`
## 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_moe_smoothquant -j
```
This will result in an executable `build/bin/tile_example_moe_smoothquant`

22
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n1024_instance.cpp Normal file
View File

@@ -0,0 +1,22 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
#if 0
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true, false>>(const S&, A);
#endif
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

13
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n1536_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n2048_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n256_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n3072_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n4096_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n4096_tp_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
// clang-format on

13
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n512_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n64_n128_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
// clang-format on

12
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n768_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

22
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n1024_instance.cpp Normal file
View File

@@ -0,0 +1,22 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
#if 0
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 8, true ,false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 4, true ,false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 2, true ,false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true ,false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true ,false>>(const S&, A);
#endif
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

13
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n1536_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n2048_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n256_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n3072_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 128, 8,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 4,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 2,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 1024, 1,true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n4096_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n4096_tp_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
// clang-format on

13
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n512_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 8, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 4, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 1, true , false>>(const S&, A);
// clang-format on

12
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n64_n128_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n768_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 4, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 12, 4, 64, 1, true , false>>(const S&, A);
// clang-format on

145
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fwd_api.cpp Normal file
View File

@@ -0,0 +1,145 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "moe_smoothquant.hpp"
template <typename DataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
ck_tile::index_t ThreadPerBlock_N_, // num threads along N
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kTwoPass_>
using trait_ = moe_smoothquant_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kTwoPass_>;
template <typename data_type>
float moe_smoothquant_dispatch(moe_smoothquant_traits /*t*/,
moe_smoothquant_args a,
const ck_tile::stream_config& s)
{
float r = -1;
// clang-format off
// rm rn tm tn vn pd 2p
if(a.hidden_size <= 64) {
r = moe_smoothquant_<trait_<data_type, 1, 1, 4, 64, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 128) {
if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 1, 4, 64, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 2, 4, 64, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 256) {
if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 1, 4, 64, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 4, 64, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 4, 4, 64, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 512) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 1, 4, 64, 8, true, false>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 4, 64, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 4, 4, 64, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 8, 4, 64, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 768) {
if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 3, 4, 64, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 6, 4, 64, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1,12, 4, 64, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 1024) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 1, 2, 128, 8, true, false>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 2, 128, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 4, 2, 128, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 4, 1, 256, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 1536) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 3, 4, 64, 8, true, false>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 3, 2, 128, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 3, 1, 256, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 6, 1, 256, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 2048) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 1, 1, 256, 8, true, false>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 1, 256, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 4, 1, 256, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 8, 1, 256, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 3072) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 3, 1, 128, 8, true, false>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 3, 1, 256, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 6, 1, 256, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 3, 1, 1024, 1, true, false>>(s, a);
}
else if(a.hidden_size <= 4096) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 1, 256, 8, true, false>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 4, 1, 256, 4, true, false>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 1, 1024, 2, true, false>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 4, 1, 1024, 1, true, false>>(s, a);
}
else if(a.hidden_size > 4096) {
if (a.hidden_size % 8 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 1, 256, 8, true, true>>(s, a);
else if (a.hidden_size % 4 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 4, 1, 256, 4, true, true>>(s, a);
else if (a.hidden_size % 2 == 0)
r = moe_smoothquant_<trait_<data_type, 1, 2, 1, 1024, 2, true, true>>(s, a);
else
r = moe_smoothquant_<trait_<data_type, 1, 4, 1, 1024, 1, true, true>>(s, a);
}
return r;
// clang-format on
}
float moe_smoothquant(moe_smoothquant_traits t,
moe_smoothquant_args a,
const ck_tile::stream_config& s)
{
if(t.data_type.compare("fp16") == 0)
{
return moe_smoothquant_dispatch<ck_tile::fp16_t>(t, a, s);
}
else if(t.data_type.compare("bf16") == 0)
{
return moe_smoothquant_dispatch<ck_tile::bf16_t>(t, a, s);
}
else
throw std::runtime_error("Without supported instances!");
}

62
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_instance_common.hpp Normal file
View File

@@ -0,0 +1,62 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "moe_smoothquant.hpp"
#include <iostream>
#pragma once
using S = ck_tile::stream_config;
using A = moe_smoothquant_args;
template <typename DataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
ck_tile::index_t ThreadPerBlock_N_, // num threads along N
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kTwoPass_>
using trait_ = moe_smoothquant_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kTwoPass_>;
template <typename Traits_>
float moe_smoothquant_(const S& s, A a)
{
using DataType = typename Traits_::DataType;
using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
typename MoeSmoothquantTypeConfig<DataType>::XDataType,
typename MoeSmoothquantTypeConfig<DataType>::XScaleDataType,
typename MoeSmoothquantTypeConfig<DataType>::ComputeDataType,
typename MoeSmoothquantTypeConfig<DataType>::YScaleDataType,
typename MoeSmoothquantTypeConfig<DataType>::QYDataType,
typename Traits_::Shape,
Traits_::kPadN,
Traits_::kTwoPass>;
using OnePassPipeline = ck_tile::SmoothquantPipelineOnePass<PipelineProblem>;
using TwoPassPipeline = ck_tile::SmoothquantPipelineTwoPass<PipelineProblem>;
using Pipeline = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::MoeSmoothquant<Pipeline>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
if(s.log_level_ > 0)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}

BIN
example/ck_tile/14_moe_smoothquant/misc/moe-sm.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 202 KiB

264
example/ck_tile/14_moe_smoothquant/moe_smoothquant.cpp Normal file
View File

@@ -0,0 +1,264 @@
#include "ck_tile/host.hpp"
#include "moe_smoothquant.hpp"
#include <cstring>
#include <set>
// different threshold for different dtype
template <typename DataType>
auto get_elimit()
{
double rtol = 1e-5;
double atol = 1e-5;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::bf16_t>()
{
double rtol = 1e-5;
double atol = 1e-5;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::int8_t>()
{
// due to rounding, int8 quantization might have 1 abs error
double rtol = 1;
double atol = 1;
return ck_tile::make_tuple(rtol, atol);
}
template <typename IndexType>
void topid_unique_gen(
std::vector<IndexType>& host_tensor, int tokens, int topk, int num_expert, int seed)
{
size_t total_size = topk * tokens;
std::srand(seed);
std::set<IndexType> unique_set;
IndexType current_v;
for(size_t i = 0; i < total_size; i++)
{
if(i % topk == 0)
{
unique_set.clear();
}
current_v = std::rand() % num_expert;
while(unique_set.find(current_v) != unique_set.end())
{
current_v = std::rand() % num_expert;
}
unique_set.insert(current_v);
host_tensor[i] = current_v;
}
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("t", "3328", "tokens dimension")
.insert("h", "4096", "hidden_size dimension")
.insert("e", "32", "experts")
.insert("k", "5", "topk")
.insert("stride", "-1", "stride per row, if -1 then equal to hidden_size")
.insert("v", "1", "cpu validation or not")
.insert("kname", "1", "print kernel name or not")
.insert("prec", "fp16", "precision")
.insert("warmup", "5", "cold iter")
.insert("repeat", "20", "hot iter");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t tokens = arg_parser.get_int("t");
ck_tile::index_t hidden_size = arg_parser.get_int("h");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = hidden_size;
ck_tile::index_t experts = arg_parser.get_int("e");
ck_tile::index_t topk = arg_parser.get_int("k");
std::string data_type = arg_parser.get_str("prec");
int kname = arg_parser.get_int("kname");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
assert(stride >= hidden_size);
using TypeConfig = MoeSmoothquantTypeConfig<DataType>;
using XDataType = typename TypeConfig::XDataType;
using XScaleDataType = typename TypeConfig::XScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({tokens, hidden_size}, {stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({experts * hidden_size});
ck_tile::HostTensor<ck_tile::index_t> topk_ids_host({tokens, topk});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({topk * tokens}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({topk * tokens}, {1});
ck_tile::HostTensor<QYDataType> qy_host_ref({topk * tokens, hidden_size}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({topk * tokens, hidden_size}, {stride, 1});
topid_unique_gen<ck_tile::index_t>(topk_ids_host.mData, tokens, topk, experts, 11937);
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem topk_ids_buf(topk_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
topk_ids_buf.ToDevice(topk_ids_host.data());
std::cout << "[" << data_type << "]"
<< " tokens:" << tokens << ", hidden_size:" << hidden_size << ", stride:" << stride
<< ", experts:" << experts << ", topk:" << topk << std::flush;
moe_smoothquant_traits traits{data_type};
moe_smoothquant_args args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
topk_ids_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
tokens,
hidden_size,
experts,
topk,
stride,
stride};
float ave_time = moe_smoothquant(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
std::size_t num_byte =
sizeof(XDataType) * tokens * hidden_size + sizeof(XScaleDataType) * topk * hidden_size +
sizeof(YScaleDataType) * topk * tokens + sizeof(QYDataType) * topk * tokens * hidden_size;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
bool pass = true;
if(do_validation)
{
using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({topk * tokens, hidden_size}, {stride, 1});
// smooth outlier
{
auto f = [&](auto i_token) {
for(int i_topk = 0; i_topk < topk; i_topk++)
{
auto i_expert = topk_ids_host(i_token, i_topk);
for(int i_h = 0; i_h < hidden_size; ++i_h)
{
auto v_xscale = ck_tile::type_convert<ComputeDataType>(
xscale_host(i_expert * hidden_size + i_h));
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(i_token, i_h));
// y_host(i_token * topk + i_topk, i_h) = v_x * v_xscale;
y_host(i_topk * tokens + i_token, i_h) = v_x * v_xscale;
}
}
};
ck_tile::make_ParallelTensorFunctor(f, tokens)(std::thread::hardware_concurrency());
}
// yscale
{
ck_tile::HostTensor<YDataType> y_rowwise_amax_host({topk * tokens});
using ReduceAmax = ck_tile::ReduceOp::AbsMax;
ck_tile::reference_reduce<ComputeDataType, ComputeDataType, YDataType>(
y_host, y_rowwise_amax_host, ReduceAmax{});
auto op = [](const auto& v0) {
return v0 /
ck_tile::type_convert<ComputeDataType>(ck_tile::numeric<QYDataType>::max());
};
ck_tile::reference_unary_elementwise<YDataType, YScaleDataType, ComputeDataType>(
y_rowwise_amax_host, yscale_host_ref, op);
yscale_buf.FromDevice(yscale_host_dev.mData.data());
auto [rtol, atol] = get_elimit<YScaleDataType>();
pass &= ck_tile::check_err(yscale_host_dev,
yscale_host_ref,
std::string("yscale Error: Incorrect results!"),
rtol,
atol);
}
// rowwise quantization
{
ck_tile::reference_rowwise_quantization2d<YDataType, YScaleDataType, QYDataType>(
y_host, yscale_host_ref, qy_host_ref);
qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == hidden_size)
{
pass = ck_tile::check_err(qy_host_dev,
qy_host_ref,
std::string("qy Error: Incorrect results!"),
rtol,
atol);
}
else
{
for(int i_r = 0; i_r < topk * tokens; i_r++)
{
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
qy_host_dev.begin() + i_r * stride +
hidden_size);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
qy_host_ref.begin() + i_r * stride +
hidden_size);
pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) +
std::string("] Error: Incorrect results!"),
rtol,
atol);
}
}
}
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
}
return pass;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16")
{
return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
}
else if(data_type == "bf16")
{
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
}
return -3;
}

114
example/ck_tile/14_moe_smoothquant/moe_smoothquant.hpp Normal file
View File

@@ -0,0 +1,114 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, 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/smoothquant.hpp"
#include <string>
template <typename DataType>
struct MoeSmoothquantTypeConfig;
template <>
struct MoeSmoothquantTypeConfig<ck_tile::half_t>
{
using XDataType = ck_tile::half_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
template <>
struct MoeSmoothquantTypeConfig<ck_tile::bf16_t>
{
using XDataType = ck_tile::bf16_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
// runtime args
struct moe_smoothquant_args : public ck_tile::MoeSmoothquantHostArgs
{
};
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
template <typename DataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
ck_tile::index_t ThreadPerBlock_N_, // num threads along N
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kTwoPass_>
struct moe_smoothquant_traits_
{
using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
static constexpr ck_tile::index_t total_warps =
(ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
// num of warps along m
static constexpr ck_tile::index_t BlockWarps_M = []() {
if constexpr(is_warp_per_row)
{
static_assert(warpSize % ThreadPerBlock_N_ == 0);
return total_warps * (warpSize / ThreadPerBlock_N_);
}
else
{
// static_assert(warpSize % ThreadPerBlock_M_ == 0);
return total_warps / (ThreadPerBlock_N_ / warpSize);
}
}();
// num of warps along n
static constexpr ck_tile::index_t BlockWarps_N = []() {
if constexpr(is_warp_per_row)
{
static_assert(warpSize % ThreadPerBlock_N_ == 0);
return 1;
}
else
{
static_assert(ThreadPerBlock_N_ % warpSize == 0);
return ThreadPerBlock_N_ / warpSize;
}
}();
static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
using BlockTile = ck_tile::sequence<Block_M, Block_N>;
using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
using WarpTile = ck_tile::sequence<Warp_M, Warp_N>;
using Vector = ck_tile::sequence<1, Vector_N_>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
static constexpr bool kPadN = kPadN_;
static constexpr bool kTwoPass = kTwoPass_;
};
template <typename Traits_>
float moe_smoothquant_(const ck_tile::stream_config& s, moe_smoothquant_args a);
// This is the public API, will be generated by script
struct moe_smoothquant_traits
{
std::string data_type;
};
float moe_smoothquant(moe_smoothquant_traits, moe_smoothquant_args, const ck_tile::stream_config&);

Some files were not shown because too many files have changed in this diff Show More