Skip to content

GitLab

Unverified Commit c87aa6c8 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge branch 'develop' into codegen_hiprtc

parents 60afb522 b70f367f
Show whitespace changes
Inline Side-by-side
Showing with 387 additions and 104 deletions
.github/CODEOWNERS
View file @ c87aa6c8
* @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
CMakeLists.txt
View file @ c87aa6c8
......@@ -183,12 +183,14 @@ message("Building CK for the following targets: ${SUPPORTED_GPU_TARGETS}")
if (SUPPORTED_GPU_TARGETS MATCHES "gfx9")
message("Enabling XDL instances")
add_definitions(-DCK_USE_XDL)
set(CK_USE_XDL "ON")
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx94")
message("Enabling FP8 gemms in ckProfiler")
add_definitions(-DCK_USE_GFX94)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
message("Enabling WMMA instances")
add_definitions(-DCK_USE_WMMA)
set(CK_USE_WMMA "ON")
endif()
option(CK_USE_FP8_ON_UNSUPPORTED_ARCH "Enable FP8 GEMM instances on older architectures" OFF)
if(CK_USE_FP8_ON_UNSUPPORTED_ARCH AND (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx908"))
......@@ -221,7 +223,7 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
endif()
set(check-coerce)
check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce)
if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000)
if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding the amdgpu-coerce-illegal-types=1")
add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
endif()
......
Dockerfile
View file @ c87aa6c8
......@@ -4,18 +4,14 @@ ARG ROCMVERSION=6.2
ARG compiler_version=""
ARG compiler_commit=""
ARG CK_SCCACHE=""
RUN set -xe
ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$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" && \
......@@ -24,14 +20,14 @@ RUN if [ "$ROCMVERSION" != "6.3" ]; then \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "6.3" ] && [ "$compiler_version" = "rc1" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.3.0.1-20.04-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.3.0.1-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.3.0.1 rel-5 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2033700; \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.3-20.04-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.3-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.3 rel-20 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2074281; \
fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
RUN amdgpu-install -y --usecase=rocm --no-dkms
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" && \
amdgpu-install -y --usecase=rocm --no-dkms
## Sccache binary built from source for ROCm, only install if CK_SCCACHE is defined
ARG SCCACHE_REPO_URL=http://compute-artifactory.amd.com/artifactory/rocm-generic-experimental/rocm-sccache
......@@ -76,66 +72,49 @@ 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 && \
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 = '$compiler_commit'"
ARG DISABLE_CACHE=0
RUN sh -c "echo compiler version = '$compiler_version'" && \
sh -c "echo compiler commit = '$compiler_commit'"
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline-open" ] ) && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
......@@ -152,9 +131,3 @@ RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd
make -j 8 ; \
else echo "using the release compiler"; \
fi
#clean-up the deb package
RUN sh -c "rm -rf amdgpu-install*"
#ENV HIP_CLANG_PATH='/llvm-project/build/bin'
#RUN sh -c "echo HIP_CLANG_PATH = '$HIP_CLANG_PATH'"
Jenkinsfile
View file @ c87aa6c8
......@@ -132,7 +132,7 @@ def buildDocker(install_prefix){
checkout scm
def image_name = getDockerImageName()
echo "Building Docker for ${image_name}"
def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg 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 ${params.COMPILER_VERSION}' "
def dockerArgs = "--squash --build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${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 "
}
......
README.md
View file @ c87aa6c8
......@@ -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
......
client_example/24_grouped_conv_activation/CMakeLists.txt
View file @ c87aa6c8
......@@ -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)
......
client_example/CMakeLists.txt
View file @ c87aa6c8
......@@ -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()
......
docs/sphinx/requirements.in
View file @ c87aa6c8
rocm-docs-core==1.8.3
rocm-docs-core==1.9.0
sphinxcontrib-bibtex==2.6.3
docs/sphinx/requirements.txt
View file @ c87aa6c8
......@@ -103,7 +103,7 @@ requests==2.32.3
# via
# pygithub
# sphinx
rocm-docs-core==1.8.3
rocm-docs-core==1.9.0
# via -r requirements.in
six==1.16.0
# via pybtex
......
example/01_gemm/CMakeLists.txt
View file @ c87aa6c8
......@@ -77,9 +77,16 @@ 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)
add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_example_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
add_example_executable(example_gemm_wmma_bf16 gemm_wmma_bf16.cpp)
add_example_dependencies(example_gemm_wmma example_gemm_wmma_bf16)
add_example_executable(example_gemm_wmma_int8 gemm_wmma_int8.cpp)
add_example_dependencies(example_gemm_wmma example_gemm_wmma_int8)
example/01_gemm/common.hpp
View file @ c87aa6c8
......@@ -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
{
......
example/01_gemm/gemm_wmma_bf16.cpp 0 → 100644
View file @ c87aa6c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp"
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = float;
using CDataType = ck::bhalf_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::MNKPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
< ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CElementOp,
GemmDefault,
1, // Prefetch stage
128, // BlockSize
64, // MPerBlock
128, // NPerBlock
64, // KPerBlock
2, // K1
16, // MPerWmma
16, // NPerWmma
2, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
4, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
2,
2,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
2,
2,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
8>;
// 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.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_wmma_int8.cpp 0 → 100644
View file @ c87aa6c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp"
using ADataType = int8_t;
using BDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = int32_t;
using CDataType = int8_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::MNKPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
< ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CElementOp,
GemmDefault,
1, // Prefetch stage
128, // BlockSize
64, // MPerBlock
128, // NPerBlock
64, // KPerBlock
2, // K1
16, // MPerWmma
16, // NPerWmma
2, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
4, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
2,
2,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
2,
2,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
8>;
// 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.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_xdl_fp16_streamk_v3.cpp
View file @ c87aa6c8
......@@ -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); }
example/01_gemm/gemm_xdl_fp8_streamk_v3.cpp 0 → 100755
View file @ c87aa6c8
// 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); }
example/01_gemm/run_gemm_example_streamk_v2.inc 100644 → 100755
View file @ c87aa6c8
......@@ -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});
......
example/01_gemm/run_gemm_example_v2.inc
View file @ c87aa6c8
......@@ -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 =
......
example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
View file @ c87aa6c8
......@@ -80,7 +80,7 @@ using RLayout = typename LayoutSettingSelector<NDimSpatial>::RLayout;
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
int init_method = 2;
bool time_kernel = false;
};
......
example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
View file @ c87aa6c8
......@@ -73,16 +73,25 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
Tensor<EDataType> conv_output_device(conv_output_g_n_k_wos_desc);
Tensor<R0DataType> r0_device(r0_desc);
std::cout << "input: " << conv_input.mDesc << std::endl;
std::cout << "weight: " << conv_weight.mDesc << std::endl;
std::cout << "output: " << conv_output_device.mDesc << std::endl;
std::cout << "reduction: " << r0_device.mDesc << std::endl << std::endl;
switch(config.init_method)
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-1, 1}(conv_weight);
break;
case 2:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input);
ck::utils::FillUniformDistribution<BDataType>{-1, 1}(conv_weight);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input);
ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight);
ck::utils::FillUniformDistribution<ADataType>{-8, 7}(conv_input);
ck::utils::FillUniformDistribution<BDataType>{-1, 1}(conv_weight);
}
DeviceMem conv_input_device_buf(sizeof(ADataType) * conv_input.mDesc.GetElementSpaceSize());
......@@ -161,15 +170,25 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
return false;
}
// XXX: DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle will not initialize r0.
r0_device_buf.SetValue(ck::NumericLimits<R0DataType>::Lowest());
const float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
if(config.time_kernel)
{
const std::size_t flop = problem_size.GetFlops();
const std::size_t num_btype = problem_size.GetByte<ADataType, BDataType, EDataType>();
const float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
const float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< conv.GetTypeString() << std::endl;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << conv.GetTypeString() << std::endl;
}
else
{
std::cout << "FINISHED: " << conv.GetTypeString() << std::endl;
}
if(config.do_verification)
{
......@@ -189,6 +208,7 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
BElementOp{},
PassThrough{});
std::cout << "\nRunning verification on CPU." << std::endl;
ref_invoker.Run(ref_argument);
Tensor<R0DataType> r0_host(r0_device.mDesc);
......@@ -273,13 +293,18 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
conv_output_device_buf.FromDevice(conv_output_device.mData.data());
r0_device_buf.FromDevice(r0_device.mData.data());
return ck::utils::check_err(conv_output_device,
auto pass = ck::utils::check_err(conv_output_device,
conv_output_host,
"Error: incorrect results! (Matrix E)",
1e-5f,
1e-4f) &&
ck::utils::check_err(
r0_device, r0_host, "Error: incorrect results! (Matrix R0)", 1e-5f, 1e-4f);
1e-3f,
1e-3f);
pass =
pass && ck::utils::check_err(
r0_device, r0_host, "Error: incorrect results! (Matrix R0)", 1e-3f, 1e-3f);
if(pass)
std::cout << "Verification on CPU: PASS" << std::endl;
return pass;
}
return true;
......
example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
View file @ c87aa6c8
......@@ -198,7 +198,7 @@ int main()
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
// init reducetion buffer to 0
// init reduction buffer to 0
r0_device_buf.SetZero();
r1_device_buf.SetZero();
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment