ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-17 17:19:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

merge upstream

Browse Source
This commit is contained in:
coderfeli
2024-12-02 05:33:33 +00:00
parent a8d88d8df8 44828b7c0f
commit 4525c5d75c
312 changed files with 18026 additions and 3981 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

105
Dockerfile
View File

@@ -4,18 +4,14 @@ ARG ROCMVERSION=6.2
ARG compiler_version=""
ARG compiler_commit=""
ARG CK_SCCACHE=""
RUN set -xe
ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
# Add rocm repository
RUN chmod 1777 /tmp
RUN apt-get update
RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
# Add rocm repository
RUN set -xe && \
useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins && \
apt-get update && apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl && \
curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN if [ "$ROCMVERSION" != "6.3" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.2.60200-1_all.deb --no-check-certificate" && \
@@ -30,8 +26,8 @@ RUN if [ "$ROCMVERSION" != "6.3" ]; then \
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
@@ -76,68 +72,47 @@ 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
# 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 +120,10 @@ 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_ub20.04_rocm6.2"
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

55
Jenkinsfile vendored
View File

@@ -32,41 +32,42 @@ 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}_${commit}"
}
}
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
else{
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_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 +132,21 @@ 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("${image_name}", dockerArgs)
withDockerRegistry([ credentialsId: "docker_test_cred", url: "" ]) {
retimage.push()
}
@@ -358,7 +363,7 @@ def buildHipClangJob(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${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')
@@ -549,7 +554,7 @@ def Build_CK(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION == "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){
@@ -737,7 +742,7 @@ def process_results(Map conf=[:]){
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
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;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''' : ""
@@ -765,7 +770,7 @@ pipeline {
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: '',

3
README.md
View File

@@ -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

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)

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);

2
client_example/CMakeLists.txt
View File

@@ -62,7 +62,7 @@ else()
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()

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
docs/sphinx/requirements.in
View File

@@ -1,2 +1,2 @@
rocm-docs-core==1.8.4
rocm-docs-core==1.9.2
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==1.8.4
rocm-docs-core==1.9.2
# 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)

2
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
{

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); }

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 =

4
example/15_grouped_gemm/grouped_gemm_multiple_d_splitk_xdl_fp16.cpp
View File

@@ -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});

2
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

4
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>
@@ -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});

4
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>
@@ -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});

4
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>
@@ -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});

18
example/15_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -168,9 +168,23 @@ 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))
{

46
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View File

@@ -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);
}}
}}
"""
@@ -614,27 +624,26 @@ 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"]):
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))
else:
assert False
return pipelines
@@ -655,9 +664,6 @@ 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,

208
example/ck_tile/01_fmha/fmha_bwd.hpp
View File

@@ -150,113 +150,113 @@ 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);
}
}();

62
example/ck_tile/01_fmha/fmha_fwd.cpp
View File

@@ -62,7 +62,7 @@ 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")
@@ -294,7 +294,8 @@ 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
@@ -321,6 +322,13 @@ 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 +364,26 @@ 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 +392,7 @@ 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(),
@@ -741,8 +753,10 @@ 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 +777,9 @@ 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 +992,9 @@ 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 +1046,7 @@ 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);

178
example/ck_tile/01_fmha/fmha_fwd.hpp
View File

@@ -165,6 +165,8 @@ 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 +175,21 @@ 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 +265,7 @@ 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 +292,87 @@ 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 +403,10 @@ 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)

6
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;
@@ -210,6 +208,8 @@ int run_gemm_example(int argc, char* argv[])
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, 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{});

53
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>;
@@ -174,8 +207,8 @@ 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})

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&);

37
example/ck_tile/14_moe_smoothquant/script/perf_test.sh Executable file
View File

@@ -0,0 +1,37 @@
EXE=build/bin/tile_example_moe_smoothquant
$EXE -t=1 -h=1 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=80 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=128 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=144 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=168 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=184 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=256 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=288 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=344 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=376 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=448 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=512 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=924 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=1024 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=1078 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=1996 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=4080 -v=1 -prec=bf16 -repeat=1000
$EXE -t=700 -h=80 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=128 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=144 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=168 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=184 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=256 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=288 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=344 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=376 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=448 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=512 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=924 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=1024 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=1078 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=1996 -v=1 -prec=fp16 -repeat=1000
$EXE -t=700 -h=4080 -v=1 -prec=fp16 -repeat=1000

30
example/ck_tile/14_moe_smoothquant/script/smoke_test.sh Executable file
View File

@@ -0,0 +1,30 @@
#!/bin/sh
EXE=build/bin/tile_example_moe_smoothquant
for pr_i in "fp16" "bf16" ; do
$EXE -prec=$pr_i -t=99 -h=13
$EXE -prec=$pr_i -t=17 -h=16
$EXE -prec=$pr_i -t=1 -h=100
$EXE -prec=$pr_i -t=4 -h=128
$EXE -prec=$pr_i -t=80 -h=127
$EXE -prec=$pr_i -t=22 -h=255 -stride=256
$EXE -prec=$pr_i -t=7 -h=599
$EXE -prec=$pr_i -t=19 -h=512
$EXE -prec=$pr_i -t=33 -h=313 -stride=1000
$EXE -prec=$pr_i -t=11 -h=510
$EXE -prec=$pr_i -t=171 -h=676 -stride=818
$EXE -prec=$pr_i -t=91 -h=636
$EXE -prec=$pr_i -t=12 -h=768 -stride=800
$EXE -prec=$pr_i -t=100 -h=766 -stride=812
$EXE -prec=$pr_i -t=31 -h=1024
$EXE -prec=$pr_i -t=64 -h=1000 -stride=1004
$EXE -prec=$pr_i -t=8 -h=1501
$EXE -prec=$pr_i -t=3 -h=1826
$EXE -prec=$pr_i -t=5 -h=2040
$EXE -prec=$pr_i -t=7 -h=2734
$EXE -prec=$pr_i -t=1 -h=3182
$EXE -prec=$pr_i -t=9 -h=4096
$EXE -prec=$pr_i -t=3 -h=8192
$EXE -prec=$pr_i -t=1 -h=10547
$EXE -prec=$pr_i -t=3 -h=17134
done

19
example/ck_tile/15_fused_moe/CMakeLists.txt Normal file
View File

@@ -0,0 +1,19 @@
set(TILE_EXAPMLE_FUSED_MOE "tile_example_fused_moe")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding ${TILE_EXAPMLE_FUSED_MOE}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${TILE_EXAPMLE_FUSED_MOE} EXCLUDE_FROM_ALL main.cpp)
target_include_directories(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${INSTANCE_SRCS})
set(TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_BUFFER_LOAD_AGPR=1) # TODO: enable load to a
list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=4) # rta
# list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1)
# list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
target_compile_options(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS})

69
example/ck_tile/15_fused_moe/README.md Normal file
View File

@@ -0,0 +1,69 @@
# fused-moe
Implementing the fused-moe block operator using ck-tile. This is a scatter/gather-group-gemm based solution, similiar to that of [vllm moe](https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_moe.py), but we introduce more kernel fusion to boost performance
![](misc/moe-0.png)
The benifit of this fused-moe:
* 1.5~2x perf boost compared with current vllm solution
* zero workspace to reduce memory footprint
* much less kernel instance, easy to maintain
# Implementation and feature support
## moe-sorting
this is a common pre-process step before the actual moe-gemm. The purpose is to transform the moe loop over from token-by-token to expert-by-expert, make sure very workgroup is working for a single expert (B matrix). Besides, we extend this op to do the zeroing of the output buffer(to be used for reduce buffer with atomic)
## moe-gemm
`moe-gemm` is a group-gemm based back-to-back gemm, where the row-id of input token comes from another buffer. Naive understanding of fused-moe is from token-by-token view as below picture:
![](misc/moe-1.png)
After `moe-sorting`, we can view this algorithm as expert-by-expert, as below:
![](misc/moe-2.png)
## optimization
summary of the key design of this fused-moe operator:
* fuse 2 group-gemm + activation + `topk-weight` multiply into single kernel, using atomic for 2nd gemm accumualation
* fuse buffer-zeroing in `moe-sorgin`, user no longer need call extra torch.zero() for the out buffer
* fused scatter-gather for row index(same as vllm)
* pre-shuffle B matric(weight) to maximize memory throughput. input(activation) keep original layout `[batch, hidden]`.
* extrem optimized pipeline using block-inline-asm(we call it `micro-kernel` or `uk`), while not breaking the *composable* design of ck
##
```
// [indexing implementation-1]
// using M_a as constexpr block_size to partition all tokens into different slices
// each slice map to one expert, and one expert can have multiple slices
// e.g. num_experts = 6, topk=3, M_a = 4, input_tokens = 5
// before sort, topk_ids is : [[0, 3, 5], [2, 3, 5], [1, 3, 5], [1, 2, 3], [1, 3, 5]]
// tok-0 tok-1 tok-2 tok-3 tok-4
// topk_weight is : [[a, b, c], [d, e, f], [g, h, i], [j, k, l], [m, n, o]] (some float number)
//
// token_id_per_expert is : [[0], [2, 3, 4], [1, 3], [0, 1, 2, 3, 4], [], [0, 1, 2, 5]]
// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5
// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]]
//
// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1)
// * this could be larger than actual, since actual tokens are on GPU
//
// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]
// |- exp-0 -|- exp-1 -|- exp-2 -|- exp-3 -|- exp-4 -|- exp-5 -|
// sorted_weight_ptr : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *, c, f, i, o]
//
// * length is max_num_tokens_padded, actual size is num_tokens_post_padded_ptr
//
// sorted_expert_ids_ptr : [0, 1, 2, 3, 3, 4, 5]
// * length is (max_num_tokens_padded + block_size - 1) / block_size
//
// num_tokens_post_padded_ptr : [28]
// num_sorted_tiles_ptr : [7]
//
// * different from vLLM
// 1) token_id stored in sorted_token_ids_ptr is actual token_id, not token_id*top_K expanded id
// 2）need sorted_weight_ptr
// 3) use num_sorted_tiles_ptr, already divided by M_a
//
// * below used for indexing
// 1) sorted_token_ids_ptr [max_num_tokens_padded]
// 2) sorted_weight_ptr
// 3) sorted_expert_ids_ptr
// 4）num_tokens_post_padded_ptr/num_sorted_tiles_ptr (select one)
//
// max_num_tokens_padded: opk_ids.numel() + num_experts * (block_size - 1)
```

52
example/ck_tile/15_fused_moe/fused_moe.hpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "fused_moesorting.hpp"
#include "fused_moegemm.hpp"
struct fused_moe_args
{
const void* a_ptr; // [m, k], input token
const void* a_scale_ptr; // [m, 1], token scale
const void* g_ptr; // [e, n, k]/[e, 2*n, k], pre-shuffle([e, nr, kr, w])
const void* d_ptr; // [e, n, k], pre-shuffle([e, nr, kr, w])
const void* g_scale_ptr; // [e, 1, n], gate(up) scale
const void* d_scale_ptr; // [e, 1, k], down scale
const void* y_smooth_scale_ptr; // [e, 1, n], smooth-quant-scale for 2nd gemm input
void* o_ptr; // [m, k], output token (no need to do zeroing)
const void* topk_ids_ptr; // [tokens, topk]
const void* topk_weight_ptr; // [tokens, topk]
void* sorted_token_ids_ptr; // [max_num_tokens_padded]
void* sorted_weight_ptr; // [max_num_tokens_padded]
void* sorted_expert_ids_ptr; // [(max_num_tokens_padded + block_size - 1) / block_size]
void* num_sorted_tiles_ptr; // [1]
ck_tile::index_t block_m; // block_m, used to devide the input
ck_tile::index_t hidden_size; // k
ck_tile::index_t intermediate_size; // n / TP, for Gate. if Gate+Up, Down need divide by 2
ck_tile::index_t num_tokens; // input number of tokens for current iteration
ck_tile::index_t num_experts; // number of groups
ck_tile::index_t topk; // need this?
ck_tile::index_t stride_token; // for input/output, stride for each row, should >= hidden_size
};
// This is the public API, will be generated by script
struct fused_moe_traits
{
std::string prec_i; // input precision
std::string prec_w; // weight precision
std::string prec_o; // output precision
std::string prec_st; // token scale data type
std::string prec_sw; // weight scale data type
std::string prec_sq; // smooth quant scale
std::string prec_kw; // topk-weight data type
int block_m;
int gate_only;
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
};
float fused_moe(fused_moe_traits, fused_moe_args, const ck_tile::stream_config&);

84
example/ck_tile/15_fused_moe/fused_moegemm.hpp Normal file
View File

@@ -0,0 +1,84 @@
// 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/fused_moe.hpp"
#include <string>
// this is only a convenient structure for creating an example
// this is not part of the host API
template <typename I, typename W, typename O, typename ST, typename SW, typename SQ, typename KW>
struct FusedMoeGemmTypeConfig;
template <typename ST, typename SW, typename SQ, typename KW>
struct FusedMoeGemmTypeConfig<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, ST, SW, SQ, KW>
{
using ADataType = ck_tile::bf16_t;
using GDataType = ck_tile::bf16_t;
using DDataType = ck_tile::bf16_t;
using AccDataType = float;
using ODataType = ck_tile::bf16_t;
using AScaleDataType = ck_tile::remove_cvref_t<ST>;
using GScaleDataType = ck_tile::remove_cvref_t<SW>;
using DScaleDataType = ck_tile::remove_cvref_t<SW>;
using YSmoothScaleDataType = ck_tile::remove_cvref_t<SQ>;
using TopkWeightDataType = ck_tile::remove_cvref_t<KW>;
using IndexDataType = ck_tile::index_t;
};
template <typename ST, typename SW, typename SQ, typename KW>
struct FusedMoeGemmTypeConfig<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, ST, SW, SQ, KW>
{
using ADataType = ck_tile::fp16_t;
using GDataType = ck_tile::fp16_t;
using DDataType = ck_tile::fp16_t;
using AccDataType = float;
using ODataType = ck_tile::fp16_t;
using AScaleDataType = ck_tile::remove_cvref_t<ST>;
using GScaleDataType = ck_tile::remove_cvref_t<SW>;
using DScaleDataType = ck_tile::remove_cvref_t<SW>;
using YSmoothScaleDataType = ck_tile::remove_cvref_t<SQ>;
using TopkWeightDataType = ck_tile::remove_cvref_t<KW>;
using IndexDataType = ck_tile::index_t;
};
template <typename ST, typename SW, typename SQ, typename KW>
struct FusedMoeGemmTypeConfig<ck_tile::int8_t, ck_tile::int8_t, ck_tile::bf16_t, ST, SW, SQ, KW>
{
using ADataType = ck_tile::int8_t;
using GDataType = ck_tile::int8_t;
using DDataType = ck_tile::int8_t;
using AccDataType = int32_t;
using ODataType = ck_tile::bf16_t;
using AScaleDataType = ck_tile::remove_cvref_t<ST>;
using GScaleDataType = ck_tile::remove_cvref_t<SW>;
using DScaleDataType = ck_tile::remove_cvref_t<SW>;
using YSmoothScaleDataType = ck_tile::remove_cvref_t<SQ>;
using TopkWeightDataType = ck_tile::remove_cvref_t<KW>;
using IndexDataType = ck_tile::index_t;
};
// runtime args
struct fused_moegemm_args : public ck_tile::FusedMoeGemmHostArgs
{
};
// This is the public API, will be generated by script
struct fused_moegemm_traits
{
std::string prec_i; // input precision
std::string prec_w; // weight precision
std::string prec_o; // output precision
std::string prec_st; // token scale data type
std::string prec_sw; // weight scale data type
std::string prec_sq; // smooth quant scale
std::string prec_kw; // topk-weight data type
int block_m;
int gate_only;
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
};
float fused_moegemm(fused_moegemm_traits, fused_moegemm_args, const ck_tile::stream_config&);

20
example/ck_tile/15_fused_moe/fused_moesorting.hpp Normal file
View File

@@ -0,0 +1,20 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/ops/fused_moe.hpp"
struct fused_moesorting_trait
{
std::string index_type;
std::string weight_type; // currently always float
};
struct fused_moesorting_args : public ck_tile::MoeSortingHostArgs
{
};
float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s);

80
example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp Normal file
View File

@@ -0,0 +1,80 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "fused_moe.hpp"
float fused_moe(fused_moe_traits t, fused_moe_args a, const ck_tile::stream_config& s)
{
auto s_sub = ck_tile::stream_config{s.stream_id_, false, s.log_level_, 0, 1};
auto o_data_bytes = [&]() {
if(t.prec_o == "fp32")
return 4;
else if(t.prec_o == "fp16" || t.prec_o == "bf16")
return 2;
else if(t.prec_o == "int8" || t.prec_o == "fp8")
return 1;
return 1;
}();
auto t0 = fused_moesorting_trait{"int32", "fp32"};
auto a0 = fused_moesorting_args{
a.topk_ids_ptr, // const void* p_topk_ids;
a.topk_weight_ptr, // const void* p_weights;
a.sorted_token_ids_ptr, // void* p_sorted_token_ids;
a.sorted_weight_ptr, // void* p_sorted_weights;
a.sorted_expert_ids_ptr, // void* p_sorted_expert_ids;
a.num_sorted_tiles_ptr, // void* p_total_tokens_post_pad;
a.o_ptr, // void* p_moe_buf;
a.num_tokens, // index_t tokens;
a.block_m, // index_t unit_size;
a.num_experts, // index_t num_experts;
a.topk, // index_t topk;
a.num_tokens * a.stride_token * o_data_bytes // index_t moe_buf_bytes;
};
auto t1 = fused_moegemm_traits{t.prec_i,
t.prec_w,
t.prec_o,
t.prec_st,
t.prec_sw,
t.prec_sq,
t.prec_kw,
t.block_m,
t.gate_only,
t.fused_quant};
auto a1 = fused_moegemm_args{
a.a_ptr, // const void* a_ptr;
a.a_scale_ptr, // const void* a_scale_ptr;
a.g_ptr, // const void* g_ptr;
a.d_ptr, // const void* d_ptr;
a.g_scale_ptr, // const void* g_scale_ptr;
a.d_scale_ptr, // const void* d_scale_ptr;
a.y_smooth_scale_ptr, // const void* y_smooth_scale_ptr;
a.o_ptr, // void* o_ptr;
a.sorted_token_ids_ptr, // const void* sorted_token_ids_ptr;
a.sorted_weight_ptr, // const void* sorted_weight_ptr;
a.sorted_expert_ids_ptr, // const void* sorted_expert_ids_ptr;
a.num_sorted_tiles_ptr, // const void* num_sorted_tiles_ptr;
a.hidden_size, // index_t hidden_size;
a.intermediate_size, // index_t intermediate_size;
a.num_tokens, // index_t num_tokens;
a.num_experts, // index_t num_experts;
a.topk, // index_t topk;
a.stride_token // index_t stride_token;
};
float r0 = -1;
float r1 = -1;
float r = ck_tile::launch_kernel(
s,
[=, &r0](const ck_tile::stream_config&) { r0 = fused_moesorting(t0, a0, s_sub); },
[=, &r1](const ck_tile::stream_config&) { r1 = fused_moegemm(t1, a1, s_sub); });
// keep unsupported case return negative
if(r0 < 0 || r1 < 0)
return -1;
return r;
}

33
example/ck_tile/15_fused_moe/instances/fused_moegemm_api.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "fused_moegemm.hpp"
#include "fused_moegemm_api_traits.hpp"
// Note: this internal API only declare, not define here, otherwise will block `make -j`
template <typename Traits_>
float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a);
template <ck_tile::index_t... Is>
using S = ck_tile::sequence<Is...>;
float fused_moegemm(fused_moegemm_traits t, fused_moegemm_args a, const ck_tile::stream_config& s)
{
// clang-format off
float r = -1;
if(t.prec_i == "bf16" && t.prec_w == "bf16" && t.prec_o == "bf16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1)
{
using t_ = fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "fp16" && t.prec_w == "fp16" && t.prec_o == "fp16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1)
{
using t_ = fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>;
r = fused_moegemm_<t_>(s, a);
}
// clang-format on
return r;
}

60
example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp Normal file
View File

@@ -0,0 +1,60 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "fused_moegemm_api_traits.hpp"
#include "ck_tile/ops/fused_moe.hpp"
#include <iostream>
template <ck_tile::index_t... Is>
using S = ck_tile::sequence<Is...>;
// do not the define of this tepmlate function inside the _api.cpp, otherwise will block make -j
template <typename Ts_>
float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a)
{
using f_traits = ck_tile::FusedMoeGemmTraits<Ts_::GateOnly, Ts_::FusedQuant == 1, 1 /*atomic*/>;
using f_shape = ck_tile::FusedMoeGemmShape<typename Ts_::BlockTile_0,
typename Ts_::WarpPerBlock_0,
typename Ts_::WarpTile_0,
typename Ts_::BlockTile_1,
typename Ts_::WarpPerBlock_0,
typename Ts_::WarpTile_0>;
using f_problem =
ck_tile::FusedMoeGemmPipelineProblem<typename Ts_::ADataType,
typename Ts_::GDataType,
typename Ts_::DDataType,
typename Ts_::AccDataType,
typename Ts_::ODataType,
typename Ts_::AScaleDataType,
typename Ts_::GScaleDataType,
typename Ts_::DScaleDataType,
typename Ts_::YSmoothScaleDataType,
typename Ts_::TopkWeightDataType,
typename Ts_::IndexDataType,
ck_tile::element_wise::FastGeluAsm, // TODO: hardcoded
f_shape,
f_traits>;
// using f_pipeline = ck_tile::FusedMoeGemmPipeline_FlatmmEx<f_problem>;
using f_pipeline = ck_tile::FusedMoeGemmPipeline_FlatmmUk<f_problem>;
using f_partitioner = ck_tile::FusedMoeGemmTilePartitioner_Linear<f_shape>;
using f_kernel = ck_tile::FusedMoeGemmKernel<f_partitioner, f_pipeline, void>;
const dim3 grids = f_kernel::GridSize(a);
constexpr dim3 blocks = f_kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
static int printed = 0;
auto kargs = f_kernel::MakeKargs(a);
if(s.log_level_ > 0 && printed == 0)
{
std::cout << ", " << f_kernel::GetName() << std::flush;
printed = 1;
}
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(f_kernel{}, grids, blocks, 0, kargs));
}

53
example/ck_tile/15_fused_moe/instances/fused_moegemm_api_traits.hpp Normal file
View File

@@ -0,0 +1,53 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <ck_tile/core.hpp>
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
template <typename I,
typename W,
typename O,
typename ST,
typename SW,
typename SQ,
typename KW,
typename BlockTIle_, // seq<b_token, b_interm, b_hidden, b_down>
typename WarpPerBlock_,
typename WarpTile_, // seq<*,*,*>, used to select mfma
ck_tile::index_t GateOnly_ = 0,
ck_tile::index_t FusedQuant_ = 0>
struct fmoe_ // traits, ugly name, only used for internal
{
using TypeConfig = FusedMoeGemmTypeConfig<I, W, O, ST, SW, SQ, KW>;
using ADataType = ck_tile::remove_cvref_t<typename TypeConfig::ADataType>;
using GDataType = ck_tile::remove_cvref_t<typename TypeConfig::GDataType>;
using DDataType = ck_tile::remove_cvref_t<typename TypeConfig::DDataType>;
using AccDataType = ck_tile::remove_cvref_t<typename TypeConfig::AccDataType>;
using ODataType = ck_tile::remove_cvref_t<typename TypeConfig::ODataType>;
using AScaleDataType = ck_tile::remove_cvref_t<typename TypeConfig::AScaleDataType>;
using GScaleDataType = ck_tile::remove_cvref_t<typename TypeConfig::GScaleDataType>;
using DScaleDataType = ck_tile::remove_cvref_t<typename TypeConfig::DScaleDataType>;
using YSmoothScaleDataType = ck_tile::remove_cvref_t<typename TypeConfig::YSmoothScaleDataType>;
using TopkWeightDataType = ck_tile::remove_cvref_t<typename TypeConfig::TopkWeightDataType>;
using IndexDataType = ck_tile::remove_cvref_t<typename TypeConfig::IndexDataType>;
static constexpr ck_tile::index_t BT_ = BlockTIle_::at(ck_tile::number<0>{}); // block token
static constexpr ck_tile::index_t BI_ =
BlockTIle_::at(ck_tile::number<1>{}); // block intermediate
static constexpr ck_tile::index_t BH_ = BlockTIle_::at(ck_tile::number<2>{}); // block hidden
static constexpr ck_tile::index_t BD_ = BlockTIle_::at(ck_tile::number<3>{}); // block down
using BlockTile_0 = ck_tile::sequence<BT_, BI_, BH_>;
using WarpPerBlock_0 = ck_tile::remove_cvref_t<WarpPerBlock_>;
using WarpTile_0 = ck_tile::remove_cvref_t<WarpTile_>;
using BlockTile_1 = ck_tile::sequence<BT_, BD_, BI_ / (GateOnly_ ? 1 : 2)>;
using WarpPerBlock_1 = ck_tile::remove_cvref_t<WarpPerBlock_>;
using WarpTile_1 = ck_tile::remove_cvref_t<WarpTile_>;
static constexpr ck_tile::index_t GateOnly = GateOnly_;
static constexpr ck_tile::index_t FusedQuant = FusedQuant_;
};

14
example/ck_tile/15_fused_moe/instances/fused_moegemm_bf16_m32.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "fused_moegemm.hpp"
#include "fused_moegemm_api_traits.hpp"
#include "fused_moegemm_api_internal.hpp"
// clang-format off
template float fused_moegemm_<
fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
// clang-format on

14
example/ck_tile/15_fused_moe/instances/fused_moegemm_fp16_m32.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "fused_moegemm.hpp"
#include "fused_moegemm_api_traits.hpp"
#include "fused_moegemm_api_internal.hpp"
// clang-format off
template float fused_moegemm_<
fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
// clang-format on

73
example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp Normal file
View File

@@ -0,0 +1,73 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "fused_moesorting.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time;
float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s)
{
if(t.weight_type == "fp32" && t.index_type == "int32")
{
if(a.num_experts > 127)
{
printf("lds size exceed, only support experts <127 \n");
return -1;
}
if(a.moe_buf_bytes % 16)
{
printf("buf set size %d unaligned, must be multiple of 16\n", a.moe_buf_bytes);
return -1;
}
using index_t = ck_tile::index_t;
using ms_weight_type = float;
index_t smem_io_unroll_num = ck_tile::integer_divide_ceil(a.tokens * a.topk, 64);
switch(smem_io_unroll_num)
{
case(1): {
MOE_SORTING_DISPATCH(1);
}
case(2): {
MOE_SORTING_DISPATCH(2);
}
case(3): {
MOE_SORTING_DISPATCH(3);
}
case(5): {
MOE_SORTING_DISPATCH(5);
}
case(6): {
MOE_SORTING_DISPATCH(6);
}
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): {
MOE_SORTING_DISPATCH(8);
}
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): {
MOE_SORTING_DISPATCH(10);
}
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: {
MOE_SORTING_DISPATCH(4);
}
}
}
return -1;
}

603
example/ck_tile/15_fused_moe/main.cpp Normal file
View File

@@ -0,0 +1,603 @@
#include <algorithm>
#include <cstring>
#include <unordered_set>
#include <vector>
#include <set>
#include "ck_tile/host.hpp"
#include "fused_moe.hpp"
// different threshold for different dtype
template <typename DataType>
auto get_elimit()
{
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::bf16_t>()
{
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
// mfma_type, 0:32x32, 1:16x16
// TODO: padding?
template <typename T>
auto shuffle_moe_weight(const ck_tile::HostTensor<T>& t, std::string mfma_dtype, int mfma_type = 0)
{
assert(t.get_lengths().size() == 3);
int b_ = t.get_lengths()[0];
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[2];
if((mfma_dtype == "bf16" || mfma_dtype == "fp16") && mfma_type == 0)
{
ck_tile::HostTensor<T> t_view({b_, n_ / 32, 32, k_ / 16, 2, 8});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5});
}
else if((mfma_dtype == "bf16" || mfma_dtype == "fp16") && mfma_type == 1)
{
ck_tile::HostTensor<T> t_view({b_, n_ / 16, 16, k_ / 32, 4, 8});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5});
}
else if((mfma_dtype == "int8" || mfma_dtype == "fp8") && mfma_type == 0)
{
ck_tile::HostTensor<T> t_view({b_, n_ / 32, 32, k_ / 32, 2, 16});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5});
}
else if((mfma_dtype == "int8" || mfma_dtype == "fp8") && mfma_type == 1)
{
ck_tile::HostTensor<T> t_view({b_, n_ / 16, 16, k_ / 64, 4, 16});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5});
}
return t;
}
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", "128", "num input tokens")
.insert("e", "32", "num of experts")
.insert("k", "5", "topk")
.insert("h", "8192", "hidden_size of this model")
.insert("i", "8192", "intermediate_size between 2 gemms of FFN")
.insert("stride", "-1", "stride per row, if -1 then equal to hidden_size")
.insert("bm", "32", "blocking factor for sorted tokens")
.insert("tp", "8", "tensor parallel size")
.insert("v", "1", "cpu validation or not")
.insert("kname", "1", "print kernel name or not")
.insert("prec_i", "bf16", "input precision")
.insert("prec_w", "bf16", "weight precision")
.insert("prec_o", "bf16", "output precision")
.insert("prec_st", "auto", "token scale data type. auto will set to fp32")
.insert("prec_sw", "auto", "weight scale data type. auto will set to fp32")
.insert("prec_sq", "auto", "(dynamic) smooth quant data type. auto will set to fp32")
.insert("prec_kw", "auto", "topk-weight data type. auto will set to fp32")
.insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant")
.insert(
"gate_only", "1", "w0(gate/up) style, 0:gate+up will double interm size, 1:only gate")
.insert("api", "0", "benchmark api set: 0:fused-moe(moe-gemm+moe-sorting), 1:moe-gemm")
.insert("balance",
"0",
"if set to 1, will try balance the expert in topk-ids(convenient for testing)")
.insert("init",
"2",
"init method. 0:random stepped float(fast). 1: random uniform, 2:rand normalized"
"normalized(slow)")
.insert("seed", "11939", "seed used to do random")
.insert("warmup", "5", "cold iter")
.insert("repeat", "20", "hot iter");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
// I:input-type, W:weight-type, O:output-type, ST:toke-scale-tpye, SW:weight-scale-type,
// SQ:smooth-quant-type, KW:topk-weight-type
template <typename I, typename W, typename O, typename ST, typename SW, typename SQ, typename KW>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t tokens = arg_parser.get_int("t");
ck_tile::index_t experts = arg_parser.get_int("e");
ck_tile::index_t topk = arg_parser.get_int("k");
ck_tile::index_t hidden_size = arg_parser.get_int("h");
ck_tile::index_t intermediate_size = arg_parser.get_int("i");
ck_tile::index_t stride = arg_parser.get_int("stride");
ck_tile::index_t block_m = arg_parser.get_int("bm");
if(stride < 0)
stride = hidden_size;
std::string prec_i = arg_parser.get_str("prec_i");
std::string prec_w = arg_parser.get_str("prec_w");
std::string prec_o = arg_parser.get_str("prec_o");
std::string prec_st = arg_parser.get_str("prec_st");
std::string prec_sw = arg_parser.get_str("prec_sw");
std::string prec_sq = arg_parser.get_str("prec_sq");
std::string prec_kw = arg_parser.get_str("prec_kw");
prec_st = (prec_st == "auto") ? "fp32" : prec_st;
prec_sw = (prec_sw == "auto") ? "fp32" : prec_sw;
prec_sq = (prec_sq == "auto") ? "fp32" : prec_sq;
prec_kw = (prec_kw == "auto") ? "fp32" : prec_kw;
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");
int fused_quant = arg_parser.get_int("fquant");
int gate_only = arg_parser.get_int("gate_only");
int api = arg_parser.get_int("api");
int balance = arg_parser.get_int("balance");
int tp = arg_parser.get_int("tp");
int init = arg_parser.get_int("init");
uint32_t seed = arg_parser.get_uint32("seed");
// w0 (Gate+Up or Gate only, N size)
ck_tile::index_t shared_intermediate_size_0 = intermediate_size * (gate_only ? 1 : 2) / tp;
// w1 (Down, N size)
ck_tile::index_t shared_intermediate_size_1 = intermediate_size / tp;
auto prec_str = [&]() {
auto base_str = prec_i;
if(prec_i != prec_w)
base_str += "x" + prec_w;
if(prec_i != prec_o)
base_str += "=" + prec_o;
if(fused_quant != 0)
{
base_str += std::string("(") + prec_st + "|" + prec_sw + "|" + prec_sq + ")";
}
return base_str;
}();
auto api_str = [&]() {
if(api == 0)
return std::string("fmoe");
else if(api == 1)
return std::string("moeg");
else if(api == 2)
return std::string("moes");
return std::string("");
}();
auto stride_str = [&]() {
if(stride == hidden_size)
return std::string("");
else
return std::string(", st:") + std::to_string(stride);
}();
std::cout << "[" << api_str << "|" << prec_str << "]"
<< " t:" << tokens << ", e:" << experts << ", k:" << topk << stride_str
<< ", hidden:" << hidden_size << ", interm:" << intermediate_size << ", tp:" << tp
<< ", shrd_interm:" << shared_intermediate_size_0 << "|" << shared_intermediate_size_1
<< ", go:" << gate_only << ", q:" << fused_quant << std::flush;
using TypeConfig = FusedMoeGemmTypeConfig<I, W, O, ST, SW, SQ, KW>;
using ADataType = typename TypeConfig::ADataType;
using GDataType = typename TypeConfig::GDataType;
using DDataType = typename TypeConfig::DDataType;
using AccDataType = typename TypeConfig::AccDataType;
using ODataType = typename TypeConfig::ODataType;
using AScaleDataType = typename TypeConfig::AScaleDataType;
using GScaleDataType = typename TypeConfig::GScaleDataType;
using DScaleDataType = typename TypeConfig::DScaleDataType;
using YSmoothScaleDataType = typename TypeConfig::YSmoothScaleDataType;
using TopkWeightDataType = typename TypeConfig::TopkWeightDataType;
using IndexDataType = typename TypeConfig::IndexDataType;
// host verify
ck_tile::HostTensor<ADataType> a_host({tokens, hidden_size}, {stride, 1});
ck_tile::HostTensor<GDataType> g_host({experts, shared_intermediate_size_0, hidden_size});
ck_tile::HostTensor<DDataType> d_host({experts, hidden_size, shared_intermediate_size_1});
ck_tile::HostTensor<ODataType> o_host({tokens, hidden_size}, {stride, 1});
ck_tile::HostTensor<AScaleDataType> sa_host({tokens});
ck_tile::HostTensor<GScaleDataType> sg_host({shared_intermediate_size_0});
ck_tile::HostTensor<DScaleDataType> sd_host({shared_intermediate_size_1});
ck_tile::HostTensor<YSmoothScaleDataType> sy_host({shared_intermediate_size_1}); // smooth-quant
ck_tile::HostTensor<IndexDataType> topk_ids_host({tokens, topk}); // to be sort
ck_tile::HostTensor<TopkWeightDataType> topk_weight_host({tokens, topk}); // to be sort
int max_num_tokens_padded = topk * tokens + experts * block_m - topk;
ck_tile::HostTensor<IndexDataType> sorted_token_ids_host({max_num_tokens_padded});
ck_tile::HostTensor<TopkWeightDataType> sorted_weight_host({max_num_tokens_padded});
ck_tile::HostTensor<IndexDataType> sorted_expert_ids_host(
{(max_num_tokens_padded + block_m - 1) / block_m});
ck_tile::HostTensor<IndexDataType> num_sorted_tiles_host({1});
if(init == 0)
{
ck_tile::FillStepRange<ADataType>{-.5f, .5f, 0.01f}(a_host);
ck_tile::FillStepRange<GDataType>{-.5f, .5f, 0.01f}(g_host);
ck_tile::FillStepRange<DDataType, false>{.5f, -.5f, -0.01f}(d_host);
ck_tile::FillStepRange<AScaleDataType>{0.f, 1.f, 0.01f}(sa_host);
ck_tile::FillStepRange<GScaleDataType>{0.f, 1.f, 0.01f}(sg_host);
ck_tile::FillStepRange<DScaleDataType>{0.f, 1.f, 0.01f}(sd_host);
ck_tile::FillStepRange<YSmoothScaleDataType>{0.f, 1.f, 0.01f}(sy_host);
ck_tile::FillStepRange<TopkWeightDataType>{-.5f, .5f, 0.01f}(topk_weight_host);
}
else if(init == 1)
{
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f, seed, true}(a_host);
ck_tile::FillUniformDistribution<GDataType>{-.5f, .5f, seed, true}(g_host);
ck_tile::FillUniformDistribution<DDataType>{-.5f, .5f, seed, true}(d_host);
ck_tile::FillUniformDistribution<AScaleDataType>{-.5f, .5f, seed, true}(sa_host);
ck_tile::FillUniformDistribution<GScaleDataType>{-.5f, .5f, seed, true}(sg_host);
ck_tile::FillUniformDistribution<DScaleDataType>{-.5f, .5f, seed, true}(sd_host);
ck_tile::FillUniformDistribution<YSmoothScaleDataType>{-.5f, .5f, seed, true}(sy_host);
ck_tile::FillUniformDistribution<TopkWeightDataType>{-.5f, .5f, seed, true}(
topk_weight_host);
}
else if(init == 2)
{
ck_tile::FillNormalDistribution<ADataType>{0.f, 1.f, seed, true}(a_host);
ck_tile::FillNormalDistribution<GDataType>{0.f, 1.f, seed, true}(g_host);
ck_tile::FillNormalDistribution<DDataType>{0.f, 1.f, seed, true}(d_host);
ck_tile::FillNormalDistribution<AScaleDataType>{0.f, 1.f, seed, true}(sa_host);
ck_tile::FillNormalDistribution<GScaleDataType>{0.f, 1.f, seed, true}(sg_host);
ck_tile::FillNormalDistribution<DScaleDataType>{0.f, 1.f, seed, true}(sd_host);
ck_tile::FillNormalDistribution<YSmoothScaleDataType>{0.f, 1.f, seed, true}(sy_host);
ck_tile::FillNormalDistribution<TopkWeightDataType>{0.f, 1.f, seed, true}(topk_weight_host);
}
// permute weight
ck_tile::HostTensor<GDataType> g_perm_host = shuffle_moe_weight(g_host, prec_w, 1);
ck_tile::HostTensor<DDataType> d_perm_host = shuffle_moe_weight(d_host, prec_w, 1);
// do moe sorting
if(balance)
{
int e_cnt = 0;
for(int i = 0; i < static_cast<int>(topk_ids_host.mData.size()); i++)
{
topk_ids_host.mData[i] = e_cnt;
e_cnt++;
if(e_cnt >= experts)
e_cnt = 0;
}
}
else
{
topid_unique_gen<IndexDataType>(topk_ids_host.mData, tokens, topk, experts, 11913);
}
// leave it here for future debug purpose
#if 0
a_host.loadtxt("../../ater/input_torch.txt");
topk_ids_host.loadtxt("../../ater/topk_ids_torch.txt", "int");
// topk_ids_host.savetxt("topk_ids_2.txt");
topk_weight_host.loadtxt("../../ater/topk_weights_torch.txt", "float");
std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl;
g_host.loadtxt("../../ater/w1_torch.txt", "float");
std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl;
d_host.loadtxt("../../ater/w2_torch.txt", "float");
std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl;
ck_tile::HostTensor<GDataType> g_perm_host = shuffle_moe_weight(g_host, prec_w, 1);
std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl;
ck_tile::HostTensor<DDataType> d_perm_host = shuffle_moe_weight(d_host, prec_w, 1);
std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl;
#endif
#if 0
std::cout << "sorted_token_ids_host:" << sorted_token_ids_host << std::endl;
std::cout << "num_sorted_tiles_host:" << num_sorted_tiles_host << std::endl;
std::cout << "sorted_expert_ids_host:" << sorted_expert_ids_host << std::endl;
std::cout << "topk_weight_host:" << topk_weight_host << std::endl;
std::cout << "sorted_weight_host:" << sorted_weight_host << std::endl;
#endif
auto cal_tflops = [&](auto ms) {
double flop_gemm_0 =
2 * static_cast<double>(tokens) * topk * shared_intermediate_size_0 * hidden_size;
double flop_gemm_1 =
2 * static_cast<double>(tokens) * topk * shared_intermediate_size_1 * hidden_size;
return (flop_gemm_0 + flop_gemm_1) / (static_cast<double>(ms) * 1e-3) / 1e12;
};
// TODO: this method we use expert-by-expert view, just for reference
auto cal_tbps = [&](auto ms) {
double token_bytes =
static_cast<double>(tokens) * topk / experts * hidden_size * sizeof(ADataType);
double w0_bytes = static_cast<double>(shared_intermediate_size_0) * experts * hidden_size *
sizeof(GDataType);
double w1_bytes = static_cast<double>(shared_intermediate_size_1) * experts * hidden_size *
sizeof(DDataType);
double o_bytes =
static_cast<double>(tokens) * topk / experts * hidden_size * sizeof(ODataType);
double topk_weights_bytes = static_cast<double>(tokens) * topk * sizeof(TopkWeightDataType);
// ignore index, they are too small
return (token_bytes + w0_bytes + w1_bytes + o_bytes + topk_weights_bytes) /
(static_cast<double>(ms) * 1e-3) / 1e12;
};
if(api == 0)
{
ck_tile::DeviceMem a_buf(a_host);
ck_tile::DeviceMem g_perm_buf(g_perm_host);
ck_tile::DeviceMem d_perm_buf(d_perm_host);
ck_tile::DeviceMem sa_buf(sa_host);
ck_tile::DeviceMem sg_buf(sg_host);
ck_tile::DeviceMem sd_buf(sd_host);
ck_tile::DeviceMem sy_buf(sy_host);
ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem topk_ids_buf(topk_ids_host);
ck_tile::DeviceMem topk_weight_buf(topk_weight_host);
ck_tile::DeviceMem sorted_token_ids_buf(
sorted_token_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_weight_buf(sorted_weight_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_expert_ids_buf(
sorted_expert_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem num_sorted_tiles_buf(
num_sorted_tiles_host.get_element_space_size_in_bytes());
fused_moe_traits traits{prec_i,
prec_w,
prec_o,
prec_st,
prec_sw,
prec_sq,
prec_kw,
block_m,
gate_only,
fused_quant};
fused_moe_args args{a_buf.GetDeviceBuffer(),
fused_quant != 0 ? sa_buf.GetDeviceBuffer() : nullptr,
g_perm_buf.GetDeviceBuffer(),
d_perm_buf.GetDeviceBuffer(),
fused_quant != 0 ? sg_buf.GetDeviceBuffer() : nullptr,
fused_quant != 0 ? sd_buf.GetDeviceBuffer() : nullptr,
fused_quant == 1 ? sy_buf.GetDeviceBuffer() : nullptr,
o_buf.GetDeviceBuffer(),
topk_ids_buf.GetDeviceBuffer(),
topk_weight_buf.GetDeviceBuffer(),
sorted_token_ids_buf.GetDeviceBuffer(),
sorted_weight_buf.GetDeviceBuffer(),
sorted_expert_ids_buf.GetDeviceBuffer(),
num_sorted_tiles_buf.GetDeviceBuffer(),
block_m,
hidden_size,
shared_intermediate_size_0,
tokens,
experts,
topk,
stride};
float ave_time = fused_moe(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
if(ave_time < 0)
{
std::cout << " not supported!" << std::endl << std::flush;
return false;
}
// float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << ", " << ave_time * 1.E3 << " us, " << cal_tflops(ave_time) << " tflops, "
<< cal_tbps(ave_time) << " TB/s" << std::flush;
bool pass = true;
if(do_validation)
{
ck_tile::reference_moe_sorting<TopkWeightDataType, IndexDataType>(
topk_ids_host,
topk_weight_host,
sorted_token_ids_host,
sorted_weight_host,
sorted_expert_ids_host,
num_sorted_tiles_host.mData[0],
experts,
block_m);
ck_tile::reference_fused_moe<AccDataType, ck_tile::element_wise::Gelu>(
a_host,
g_host,
d_host,
sa_host,
sg_host,
sd_host,
sy_host,
o_host,
sorted_token_ids_host,
sorted_weight_host,
sorted_expert_ids_host,
num_sorted_tiles_host,
topk_ids_host,
block_m,
tokens,
experts,
hidden_size,
shared_intermediate_size_0,
topk,
gate_only);
auto o_dev = o_buf.ToHost<ODataType>();
// o_dev.savetxt("gpu-out.txt", "float");
auto [rtol, atol] = get_elimit<ADataType>();
pass &= ck_tile::check_err(
o_dev, o_host, std::string("OUT Error: Incorrect results!"), rtol, atol);
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush;
}
std::cout << std::flush << std::endl;
return pass;
}
else if(api == 1)
{
ck_tile::reference_moe_sorting<TopkWeightDataType, IndexDataType>(
topk_ids_host,
topk_weight_host,
sorted_token_ids_host,
sorted_weight_host,
sorted_expert_ids_host,
num_sorted_tiles_host.mData[0],
experts,
block_m);
// done, preparing GPU buffer
ck_tile::DeviceMem a_buf(a_host);
ck_tile::DeviceMem g_perm_buf(g_perm_host);
ck_tile::DeviceMem d_perm_buf(d_perm_host);
ck_tile::DeviceMem sa_buf(sa_host);
ck_tile::DeviceMem sg_buf(sg_host);
ck_tile::DeviceMem sd_buf(sd_host);
ck_tile::DeviceMem sy_buf(sy_host);
ck_tile::DeviceMem o_buf(o_host);
// manually clear output buffer for atomic
o_buf.SetZero();
//
ck_tile::DeviceMem sorted_token_ids_buf(sorted_token_ids_host);
ck_tile::DeviceMem sorted_weight_buf(sorted_weight_host);
ck_tile::DeviceMem sorted_expert_ids_buf(sorted_expert_ids_host);
ck_tile::DeviceMem num_sorted_tiles_buf(num_sorted_tiles_host);
fused_moegemm_traits traits{prec_i,
prec_w,
prec_o,
prec_st,
prec_sw,
prec_sq,
prec_kw,
block_m,
gate_only,
fused_quant};
fused_moegemm_args args{a_buf.GetDeviceBuffer(),
fused_quant != 0 ? sa_buf.GetDeviceBuffer() : nullptr,
g_perm_buf.GetDeviceBuffer(),
d_perm_buf.GetDeviceBuffer(),
fused_quant != 0 ? sg_buf.GetDeviceBuffer() : nullptr,
fused_quant != 0 ? sd_buf.GetDeviceBuffer() : nullptr,
fused_quant == 1 ? sy_buf.GetDeviceBuffer() : nullptr,
o_buf.GetDeviceBuffer(),
sorted_token_ids_buf.GetDeviceBuffer(),
sorted_weight_buf.GetDeviceBuffer(),
sorted_expert_ids_buf.GetDeviceBuffer(),
num_sorted_tiles_buf.GetDeviceBuffer(),
hidden_size,
shared_intermediate_size_0,
tokens,
experts,
topk,
stride};
float ave_time = fused_moegemm(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
if(ave_time < 0)
{
std::cout << " not supported!" << std::endl << std::flush;
return false;
}
// float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << ", " << ave_time * 1.E3 << " us, " << cal_tflops(ave_time) << " tflops, "
<< cal_tbps(ave_time) << " TB/s" << std::flush;
bool pass = true;
if(do_validation)
{
ck_tile::reference_fused_moe<AccDataType, ck_tile::element_wise::Gelu>(
a_host,
g_host,
d_host,
sa_host,
sg_host,
sd_host,
sy_host,
o_host,
sorted_token_ids_host,
sorted_weight_host,
sorted_expert_ids_host,
num_sorted_tiles_host,
topk_ids_host,
block_m,
tokens,
experts,
hidden_size,
shared_intermediate_size_0,
topk,
gate_only);
auto o_dev = o_buf.ToHost<ODataType>();
// o_dev.savetxt("gpu-out.txt", "float");
auto [rtol, atol] = get_elimit<ADataType>();
pass &= ck_tile::check_err(
o_dev, o_host, std::string("OUT Error: Incorrect results!"), rtol, atol);
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush;
}
std::cout << std::flush << std::endl;
return pass;
}
return false;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
std::string prec_i = arg_parser.get_str("prec_i");
std::string prec_w = arg_parser.get_str("prec_w");
std::string prec_o = arg_parser.get_str("prec_o");
std::string prec_st = arg_parser.get_str("prec_st");
std::string prec_sw = arg_parser.get_str("prec_sw");
std::string prec_sq = arg_parser.get_str("prec_sq");
std::string prec_kw = arg_parser.get_str("prec_kw");
prec_st = (prec_st == "auto") ? "fp32" : prec_st;
prec_sw = (prec_sw == "auto") ? "fp32" : prec_sw;
prec_sq = (prec_sq == "auto") ? "fp32" : prec_sq;
prec_kw = (prec_kw == "auto") ? "fp32" : prec_kw;
// no dynamic quant case
if(prec_i == "bf16" && prec_w == "bf16" && prec_o == "bf16" && prec_kw == "fp32")
{
return run<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float>(
arg_parser)
? 0
: -2;
}
else if(prec_i == "fp16" && prec_w == "fp16" && prec_o == "fp16" && prec_kw == "fp32")
{
return run<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float>(
arg_parser)
? 0
: -2;
}
return -3;
}

BIN
example/ck_tile/15_fused_moe/misc/moe-0.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 75 KiB

BIN
example/ck_tile/15_fused_moe/misc/moe-1.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 90 KiB

BIN
example/ck_tile/15_fused_moe/misc/moe-2.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 124 KiB

BIN
example/ck_tile/15_fused_moe/misc/moe-3.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 18 KiB

1
example/ck_tile/16_batched_gemm/CMakeLists.txt Normal file
View File

@@ -0,0 +1 @@
add_executable(tile_example_batched_gemm EXCLUDE_FROM_ALL batched_gemm.cpp)

37
example/ck_tile/16_batched_gemm/README.md Normal file
View File

@@ -0,0 +1,37 @@
# Batched GEMM
This folder contains example for batched GEMM using ck_tile tile-programming implementation.
## build
```
# in the root of ck_tile
mkdir build && cd build
# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
sh ../script/cmake-ck-dev.sh ../ <arch>
make tile_example_batched_gemm -j
```
This will result in an executable `build/bin/tile_example_batched_gemm`
## example
```
args:
-m m dimension (default:256)
-n n dimension (default:128)
-k k dimension (default:128)
-a_layout A tensor data layout (default:R) (R for Row, C for Col)
-b_layout B tensor data layout (default:R) (R for Row, C for Col)
-c_layout C tensor data layout (default:R) (R for Row, C for Col)
-stride_a Tensor A stride (default:128)
-stride_b Tensor B stride (default:128)
-stride_c Tensor C stride (default:128)
-batch_stride_a Batch A stride (default:32768)
-batch_stride_b Batch B stride (default:16384)
-batch_stride_c Batch C stride (default:32768)
-batch_count Batch count (default:16)
-v 0. No validation, 1. Validation on CPU, 2. Validation on GPU (default:2)
-e Absolute error tolerance (default:1e-5)
-prec data type. fp16/bf16/fp8/bf8 (default:fp16)
-warmup number of iterations before benchmark the kernel (default:10)
-repeat number of iterations to benchmark the kernel (default:100)
-timer gpu:gpu timer, cpu:cpu timer (default:gpu)
```

103
example/ck_tile/16_batched_gemm/batched_gemm.cpp Normal file
View File

@@ -0,0 +1,103 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include "ck_tile/core.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/host.hpp"
#include "batched_gemm.hpp"
template <typename ALayout, typename BLayout, typename CLayout>
float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config& s)
{
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr int kBlockPerCu = 1;
// This part comes from the Codegen
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTilePartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
0,
TilePartitioner::kM,
TilePartitioner::kN>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args);
const dim3 grids = Kernel::GridSize(args);
constexpr dim3 blocks = Kernel::BlockSize();
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }

63
example/ck_tile/16_batched_gemm/batched_gemm.hpp Normal file
View File

@@ -0,0 +1,63 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
template <typename DataType>
struct BatchedGemmTypeConfig;
template <>
struct BatchedGemmTypeConfig<ck_tile::half_t>
{
using ADataType = ck_tile::half_t;
using BDataType = ck_tile::half_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
using Types = BatchedGemmTypeConfig<ck_tile::half_t>;
// Specific type aliases for easy access
using ADataType = Types::ADataType;
using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType;
struct batched_gemm_kargs : public ck_tile::BatchedGemmHostArgs
{
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "256", "m dimension")
.insert("n", "128", "n dimension")
.insert("k", "128", "k dimension")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "R", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("batch_stride_a", "32768", "Batch A stride")
.insert("batch_stride_b", "16384", "Batch B stride")
.insert("batch_stride_c", "32768", "Batch C stride")
.insert("batch_count", "16", "Batch count")
.insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
// host API
float batched_gemm(batched_gemm_kargs args, const ck_tile::stream_config& s);

253
example/ck_tile/16_batched_gemm/run_batched_gemm_example.inc Normal file
View File

@@ -0,0 +1,253 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename ALayout, typename BLayout, typename CLayout>
float invoke_batched_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf,
ck_tile::DeviceMem& c_m_n_dev_buf,
ck_tile::index_t M,
ck_tile::index_t N,
ck_tile::index_t K,
ck_tile::index_t stride_A,
ck_tile::index_t stride_B,
ck_tile::index_t stride_C,
ck_tile::index_t batch_stride_A,
ck_tile::index_t batch_stride_B,
ck_tile::index_t batch_stride_C,
ck_tile::index_t batch_count,
int n_warmup,
int n_repeat)
{
batched_gemm_kargs args;
args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.M = M;
args.N = N;
args.K = K;
args.stride_A = stride_A;
args.stride_B = stride_B;
args.stride_C = stride_C;
args.batch_stride_A = batch_stride_A;
args.batch_stride_B = batch_stride_B;
args.batch_stride_C = batch_stride_C;
args.batch_count = batch_count;
float ave_time = batched_gemm<ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::string op_name{"Batched Gemm"};
std::size_t flop = std::size_t(2) * batch_count * M * N * K;
std::size_t num_byte = sizeof(ADataType) * batch_count * M * K +
sizeof(BDataType) * batch_count * N * K +
sizeof(CDataType) * batch_count * 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
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
<< " batch_stride_A =" << batch_stride_A << " batch_stride_B =" << batch_stride_B
<< " batch_stride_C =" << batch_stride_C << " batch_count =" << batch_count << " : "
<< ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
return ave_time;
}
template <typename ALayout, typename BLayout, typename CLayout>
int run_batched_gemm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
ck_tile::index_t batch_stride_A = arg_parser.get_int("batch_stride_a");
ck_tile::index_t batch_stride_B = arg_parser.get_int("batch_stride_b");
ck_tile::index_t batch_stride_C = arg_parser.get_int("batch_stride_c");
ck_tile::index_t batch_count = arg_parser.get_int("batch_count");
int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat");
using namespace ck_tile::literals;
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return ck_tile::HostTensorDescriptor({batch_count_, row, col},
{batch_stride, stride, 1_uz});
}
else
{
return ck_tile::HostTensorDescriptor({batch_count_, row, col},
{batch_stride, 1_uz, stride});
}
};
auto f_get_default_stride = [](std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(stride == 0)
{
// give a chance if stride is zero, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
stride_A = f_get_default_stride(M, K, stride_A, a_layout);
stride_B = f_get_default_stride(K, N, stride_B, b_layout);
stride_C = f_get_default_stride(M, N, stride_C, c_layout);
ck_tile::HostTensor<ADataType> a_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, a_layout));
ck_tile::HostTensor<BDataType> b_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, b_layout));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, c_layout));
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
a_m_k_dev_buf.ToDevice(a_m_k.data());
b_k_n_dev_buf.ToDevice(b_k_n.data());
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
invoke_batched_gemm<ALayout, BLayout, CLayout>(a_m_k_dev_buf,
b_k_n_dev_buf,
c_m_n_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_count,
n_warmup,
n_repeat);
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
bool pass = true;
if(arg_parser.get_int("v") == 1)
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, CLayout{}));
c_m_n_host_ref.SetZero();
const auto b_n_k = b_k_n.transpose({0, 2, 1});
ck_tile::reference_batched_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k, b_n_k, c_m_n_host_ref);
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_host_ref);
std::cout << "The CPU veification result is:" << (pass ? "correct" : "fail") << std::endl;
}
else if(arg_parser.get_int("v") == 2)
{
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, CLayout{}));
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
c_m_n_gpu_ref.SetZero();
c_m_n_gpu_buf_ref.SetZero();
ck_tile::reference_batched_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(a_m_k_dev_buf,
b_k_n_dev_buf,
c_m_n_gpu_buf_ref,
M,
N,
K,
stride_A,
stride_B,
stride_C,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_count);
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref);
std::cout << "The GPU verification result is: " << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}
int run_batched_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "R")
{
return run_batched_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
else if(a_layout == "R" && b_layout == "C")
{
return run_batched_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
// TODO: Fixme: with latest changes to GemmPipelineAGmemBGmemCRegV1DefaultPolicy below do not
// work else if(a_layout == "C" && b_layout == "C")
// {
// return run_batched_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
// }
// else if(a_layout == "C" && b_layout == "R")
// {
// return run_batched_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
// }
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}

3
example/ck_tile/CMakeLists.txt
View File

@@ -13,3 +13,6 @@ add_subdirectory(10_rmsnorm2d)
add_subdirectory(11_add_rmsnorm2d_rdquant)
add_subdirectory(12_smoothquant)
add_subdirectory(13_moe_sorting)
add_subdirectory(14_moe_smoothquant)
add_subdirectory(15_fused_moe)
add_subdirectory(16_batched_gemm)

0
library/include/ck/library/utility/algorithm.hpp → include/ck/library/utility/algorithm.hpp
View File

0
library/include/ck/library/utility/check_err.hpp → include/ck/library/utility/check_err.hpp
View File

0
library/include/ck/library/utility/conv_common.hpp → include/ck/library/utility/conv_common.hpp
View File

0
library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp → include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
View File

0
library/include/ck/library/utility/convolution_parameter.hpp → include/ck/library/utility/convolution_parameter.hpp
View File

0
library/include/ck/library/utility/device_memory.hpp → include/ck/library/utility/device_memory.hpp
View File

0
library/include/ck/library/utility/fill.hpp → include/ck/library/utility/fill.hpp
View File

0
library/include/ck/library/utility/host_common_util.hpp → include/ck/library/utility/host_common_util.hpp
View File

0
library/include/ck/library/utility/host_gemm.hpp → include/ck/library/utility/host_gemm.hpp
View File

0
library/include/ck/library/utility/host_tensor.hpp → include/ck/library/utility/host_tensor.hpp
View File

0
library/include/ck/library/utility/host_tensor_generator.hpp → include/ck/library/utility/host_tensor_generator.hpp
View File

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