Skip to content

GitLab

Unverified Commit 9f8ab221 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into add_int8_wmma_example_instance

parents 755ace59 b4fc4d0b
Hide whitespace changes
Inline Side-by-side
Showing with 908 additions and 155 deletions
CHANGELOG.md
View file @ 9f8ab221
# Change Log for Composable Kernel # Changelog for Composable Kernel
Full documentation for Composable Kernel is not yet available. Full documentation for Composable Kernel is not yet available.
## CK 0.2.0 for ROCm 5.5.0 ## (Unreleased) CK for ROCm 6.0.0
### Fixed ### Fixes
- Fixed a bug in 6-dimensional kernels (#555). - Fixed a hazard associated with inline v_dot (#808)
- Fixed grouped ConvBwdWeight test case failure (#524). - Fixed two bugs in grouped convolution backward data without K padding (#848 #876)
### Optimizations ### Optimizations
- Improve proformance of normalization kernel None
### Added ### Additions
- Added new cmake flag "DL_KERNELS" must be set to "ON" in order to build the gemm_dl and batched_gemm_multi_d_dl instances. - Added an image to a column kernel (#867)
- Added new cmake flag "DTYPES" which could be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build instance of select data types. - Added a column to an image kernel (#930)
- Added new cmake flag "INSTANCES_ONLY" which will only build CK library and instances without the tests, examples, or profiler. - Support for 3D grouped convolution on RDNA 3 GPUs (#935, #950, #985)
- Added new feature: if GPU_TARGETS is not set on cmake command line, CK will be built for all targets supported by compiler. - Grouped convolution support for small K and C (#822 #879 #897)
- Added support on MI300A/MI300X. - Support for NHWGC (2D and 3D) grouped convolution backward weight (#769 #804)
- Added support on NAVI3x. - Support for bf16/f32/f16 and NHWGC (2D and 3D) grouped convolution backward data (#757 #799)
- Added user tutorial (#563). - Support for Batched Gemm DL (#732)
- Added more instances for irregular GEMM sizes (#560).
- Added inter-wave consumer-producer programming model for GEMM kernels (#310). ### Changes
- Added multi-D GEMM client APIs (#534). - Changed the grouped convolution API to maintain consistency with other convolution kernels (#817)
- Added multi-embeddings support (#542).
- Added Navi3x blockwise GEMM and real GEMM support (#541). ## CK 0.2.0 for ROCm 5.7.0
- Added Navi grouped ConvBwdWeight support (#505).
- Added MaxPool, AvgPool forward (#815). ### Fixes
- Added MaxPool backward (#750). - Fixed a bug in 6-dimensional kernels (#555)
- Fixed a test case failure with grouped convolution backward weight (#524)
### Changed
- Changed ... ### Optimizations
- Improved the performance of the normalization kernel
### Additions
- New CMake flags:
- "DL_KERNELS"-- Must be set to "ON" in order to build the gemm_dl and batched_gemm_multi_d_dl instances
- "DTYPES" -- Can be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build an instance of the specified data types
- "INSTANCES_ONLY" -- Only builds CK library and instances without tests, examples, or profiler
- New feature: if GPU_TARGETS is not set in the CMake command line, CK will be built for all targets supported by the compiler
- Support for MI300A/MI300X
- Support for AMD RDNA 3
- New user tutorial (#563)
- Additional instances for irregular GEMM sizes (#560)
- New inter-wave consumer-producer programming model for GEMM kernels (#310)
- GEMM with support multiple elementwise fusions (multi-D) (#534)
- Multi-embeddings support (#542)
- AMD RDNA 3 blockwise GEMM and real GEMM support (#541)
- AMD RDNA grouped convolution backward weight support (#505)
- MaxPool and AvgPool forward (#815); MaxPool backward (#750)
### Changes
None
CMakeLists.txt
View file @ 9f8ab221
cmake_minimum_required(VERSION 3.14) cmake_minimum_required(VERSION 3.14)
if(POLICY CMP0140)
# policies CMP0140 not known to CMake until 3.25
cmake_policy(SET CMP0140 NEW)
endif()
# This has to be initialized before the project() command appears # This has to be initialized before the project() command appears
# Set the default of CMAKE_BUILD_TYPE to be release, unless user specifies with -D. MSVC_IDE does not use CMAKE_BUILD_TYPE # Set the default of CMAKE_BUILD_TYPE to be release, unless user specifies with -D. MSVC_IDE does not use CMAKE_BUILD_TYPE
...@@ -28,12 +32,10 @@ if (DTYPES) ...@@ -28,12 +32,10 @@ if (DTYPES)
if (DTYPES MATCHES "fp8") if (DTYPES MATCHES "fp8")
add_definitions(-DCK_ENABLE_FP8) add_definitions(-DCK_ENABLE_FP8)
set(CK_ENABLE_FP8 "ON") set(CK_ENABLE_FP8 "ON")
add_compile_options(-Wno-bit-int-extension)
endif() endif()
if (DTYPES MATCHES "bf8") if (DTYPES MATCHES "bf8")
add_definitions(-DCK_ENABLE_BF8) add_definitions(-DCK_ENABLE_BF8)
set(CK_ENABLE_BF8 "ON") set(CK_ENABLE_BF8 "ON")
add_compile_options(-Wno-bit-int-extension)
endif() endif()
if (DTYPES MATCHES "fp16") if (DTYPES MATCHES "fp16")
add_definitions(-DCK_ENABLE_FP16) add_definitions(-DCK_ENABLE_FP16)
...@@ -55,9 +57,11 @@ if (DTYPES) ...@@ -55,9 +57,11 @@ if (DTYPES)
else() else()
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16) add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_ALL_DTYPES "ON") set(CK_ENABLE_ALL_DTYPES "ON")
add_compile_options(-Wno-bit-int-extension) # enable fp8 and bf8
endif() endif()
#for f8/bf8_t type
add_compile_options(-Wno-bit-int-extension)
if(DL_KERNELS) if(DL_KERNELS)
add_definitions(-DDL_KERNELS) add_definitions(-DDL_KERNELS)
set(CK_ENABLE_DL_KERNELS "ON") set(CK_ENABLE_DL_KERNELS "ON")
...@@ -102,26 +106,30 @@ message("checking which targets are supported") ...@@ -102,26 +106,30 @@ message("checking which targets are supported")
#Setting GPU_TARGETS on command line will override this list #Setting GPU_TARGETS on command line will override this list
if(NOT PROFILER_ONLY) if(NOT PROFILER_ONLY)
rocm_check_target_ids(DEFAULT_GPU_TARGETS rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx900;gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102") TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
else() else()
add_definitions(-DPROFILER_ONLY) add_definitions(-DPROFILER_ONLY)
set(GPU_TARGETS "" CACHE STRING "" FORCE)
if(GPU_TARGETS) if(GPU_TARGETS)
message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx9, gfx10, or gfx11") message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, or gfx11")
endif() endif()
if(GPU_ARCH MATCHES "gfx9") if(GPU_ARCH MATCHES "gfx90")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx900;gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942") rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx908;gfx90a")
elseif(GPU_ARCH MATCHES "gfx94")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx940;gfx941;gfx942")
elseif(GPU_ARCH MATCHES "gfx10") elseif(GPU_ARCH MATCHES "gfx10")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030") rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030")
elseif(GPU_ARCH MATCHES "gfx11") elseif(GPU_ARCH MATCHES "gfx11")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102") rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102")
else() else()
message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx9, gfx10, or gfx11") message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, or gfx11")
endif() endif()
set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
endif() endif()
message("Supported GPU_TARGETS= ${DEFAULT_GPU_TARGETS}") message("Supported GPU_TARGETS= ${DEFAULT_GPU_TARGETS}")
set(AMDGPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " ") set(AMDGPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
if(GPU_TARGETS) if(GPU_TARGETS)
message("Building CK for the following targets: ${GPU_TARGETS}") message("Building CK for the following targets: ${GPU_TARGETS}")
...@@ -381,31 +389,31 @@ IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu ...@@ -381,31 +389,31 @@ IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu
set(cmake_instance) set(cmake_instance)
file(READ "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}/CMakeLists.txt" cmake_instance) file(READ "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}/CMakeLists.txt" cmake_instance)
set(add_inst 0) set(add_inst 0)
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8") if(("${cmake_instance}" MATCHES "fp8" OR "${cmake_instance}" MATCHES "_f8") AND DTYPES MATCHES "fp8")
#message("fp8 instance found!") #message("fp8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf8\" " AND DTYPES MATCHES "bf8") if(("${cmake_instance}" MATCHES "bf8" OR "${cmake_instance}" MATCHES "_b8") AND DTYPES MATCHES "bf8")
#message("bf8 instance found!") #message("bf8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16") if(("${cmake_instance}" MATCHES "fp16" OR "${cmake_instance}" MATCHES "_f16") AND DTYPES MATCHES "fp16")
#message("fp16 instance found!") #message("fp16 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32") if(("${cmake_instance}" MATCHES "fp32" OR "${cmake_instance}" MATCHES "_f32") AND DTYPES MATCHES "fp32")
#message("fp32 instance found!") #message("fp32 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64") if(("${cmake_instance}" MATCHES "fp64" OR "${cmake_instance}" MATCHES "_f64") AND DTYPES MATCHES "fp64")
#message("fp64 instance found!") #message("fp64 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16") if(("${cmake_instance}" MATCHES "bf16" OR "${cmake_instance}" MATCHES "_b16") AND DTYPES MATCHES "bf16")
#message("bf16 instance found!") #message("bf16 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8") if(("${cmake_instance}" MATCHES "int8" OR "${cmake_instance}" MATCHES "_i8") AND DTYPES MATCHES "int8")
#message("int8 instance found!") #message("int8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
...@@ -438,14 +446,14 @@ if(NOT DEFINED INSTANCES_ONLY) ...@@ -438,14 +446,14 @@ if(NOT DEFINED INSTANCES_ONLY)
rocm_package_setup_component(profiler rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel LIBRARY_NAME composablekernel
PACKAGE_NAME ckProfiler PACKAGE_NAME ckprofiler
) )
add_subdirectory(profiler) add_subdirectory(profiler)
else() else()
#When building PROFILER_ONLY, label the package with GPU_ARCH #When building PROFILER_ONLY, label the package with GPU_ARCH
rocm_package_setup_component(profiler rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel LIBRARY_NAME composablekernel
PACKAGE_NAME ckProfiler_${GPU_ARCH} PACKAGE_NAME ckprofiler_${GPU_ARCH}
) )
add_subdirectory(profiler) add_subdirectory(profiler)
endif() endif()
......
Dockerfile
View file @ 9f8ab221
FROM ubuntu:20.04 FROM ubuntu:20.04
ARG DEBIAN_FRONTEND=noninteractive ARG DEBIAN_FRONTEND=noninteractive
ARG ROCMVERSION=5.6 ARG ROCMVERSION=5.7
ARG compiler_version="" ARG compiler_version=""
ARG compiler_commit="" ARG compiler_commit=""
...@@ -16,25 +16,12 @@ RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl ...@@ -16,25 +16,12 @@ RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN wget https://repo.radeon.com/amdgpu-install/5.6/ubuntu/focal/amdgpu-install_5.6.50600-1_all.deb --no-check-certificate RUN wget https://repo.radeon.com/amdgpu-install/5.7/ubuntu/focal/amdgpu-install_5.7.50700-1_all.deb --no-check-certificate
RUN apt-get update && \ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_5.7.50700-1_all.deb
DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
./amdgpu-install_5.6.50600-1_all.deb RUN wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
RUN if [ "$ROCMVERSION" != "5.7" ]; then \ 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'
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "" ] || [ "$compiler_version" = "amd-stg-open" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \
apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \
amdgpu-repo --amdgpu-build=1609671 --rocm-build=compute-rocm-npi-mi300/1354; \
elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "rc1" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \
apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.7 rel-19 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=1637781; \
fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
RUN amdgpu-install -y --usecase=rocm --no-dkms RUN amdgpu-install -y --usecase=rocm --no-dkms
...@@ -62,6 +49,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -62,6 +49,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
vim \ vim \
nano \ nano \
zlib1g-dev \ zlib1g-dev \
zip \
openssh-server \ openssh-server \
clang-format-12 \ clang-format-12 \
kmod && \ kmod && \
...@@ -114,7 +102,7 @@ RUN sh -c "echo compiler commit = '$compiler_commit'" ...@@ -114,7 +102,7 @@ RUN sh -c "echo compiler commit = '$compiler_commit'"
RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \ RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \ git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && mkdir build && cd build && \ cd llvm-project && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \ cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
make -j 8 ; \ make -j 8 ; \
else echo "using the release compiler"; \ else echo "using the release compiler"; \
fi fi
...@@ -122,7 +110,7 @@ RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; ...@@ -122,7 +110,7 @@ RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ];
RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \ RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \ git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \ cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \ cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
make -j 8 ; \ make -j 8 ; \
else echo "using the release compiler"; \ else echo "using the release compiler"; \
fi fi
......
Jenkinsfile
View file @ 9f8ab221
...@@ -33,7 +33,7 @@ def runShell(String command){ ...@@ -33,7 +33,7 @@ def runShell(String command){
def getDockerImageName(){ def getDockerImageName(){
def img def img
if (params.ROCMVERSION != "5.7"){ if (params.ROCMVERSION != "6.0"){
if (params.COMPILER_VERSION == "") { if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}" img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
} }
...@@ -526,6 +526,26 @@ def Build_CK(Map conf=[:]){ ...@@ -526,6 +526,26 @@ def Build_CK(Map conf=[:]){
stash "ckprofiler_0.2.0_amd64.deb" stash "ckprofiler_0.2.0_amd64.deb"
} }
} }
if (params.hipTensor_test && navi_node == 0 ){
//build and test hipTensor
sh """#!/bin/bash
rm -rf "${params.hipTensor_branch}".zip
rm -rf hipTensor-"${params.hipTensor_branch}"
wget https://github.com/ROCmSoftwarePlatform/hipTensor/archive/refs/heads/"${params.hipTensor_branch}".zip
unzip -o "${params.hipTensor_branch}".zip
"""
dir("hipTensor-${params.hipTensor_branch}"){
sh """#!/bin/bash
mkdir -p build
ls -ltr
CC=hipcc CXX=hipcc cmake -Bbuild . -D CMAKE_PREFIX_PATH="/opt/rocm;${env.WORKSPACE}/install"
cmake --build build -- -j
"""
}
dir("hipTensor-${params.hipTensor_branch}/build"){
sh 'ctest'
}
}
} }
} }
} }
...@@ -613,8 +633,8 @@ def process_results(Map conf=[:]){ ...@@ -613,8 +633,8 @@ def process_results(Map conf=[:]){
} }
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=5.7;COMPILER_VERSION=rc1 CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=5.7;COMPILER_VERSION=
0 21 * * * % ROCMVERSION=5.6;COMPILER_VERSION=;COMPILER_COMMIT= 0 21 * * * % ROCMVERSION=5.7;COMPILER_VERSION=;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : "" 0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
pipeline { pipeline {
...@@ -632,8 +652,8 @@ pipeline { ...@@ -632,8 +652,8 @@ pipeline {
description: "Force building docker image (default: false), set to true if docker image needs to be updated.") description: "Force building docker image (default: false), set to true if docker image needs to be updated.")
string( string(
name: 'ROCMVERSION', name: 'ROCMVERSION',
defaultValue: '5.6', defaultValue: '5.7',
description: 'Specify which ROCM version to use: 5.6 (default).') description: 'Specify which ROCM version to use: 5.7 (default).')
string( string(
name: 'COMPILER_VERSION', name: 'COMPILER_VERSION',
defaultValue: '', defaultValue: '',
...@@ -654,6 +674,15 @@ pipeline { ...@@ -654,6 +674,15 @@ pipeline {
name: "DL_KERNELS", name: "DL_KERNELS",
defaultValue: false, defaultValue: false,
description: "Select whether to build DL kernels (default: OFF)") description: "Select whether to build DL kernels (default: OFF)")
booleanParam(
name: "hipTensor_test",
defaultValue: true,
description: "Use the CK build to verify hipTensor build and tests (default: ON)")
string(
name: 'hipTensor_branch',
defaultValue: 'mainline',
description: 'Specify which branch of hipTensor to use (default: mainline)')
} }
environment{ environment{
dbuser = "${dbuser}" dbuser = "${dbuser}"
...@@ -713,8 +742,8 @@ pipeline { ...@@ -713,8 +742,8 @@ pipeline {
} }
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941" """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
} }
steps{ steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local') Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......
client_example/05_layernorm/layernorm2d.cpp
View file @ 9f8ab221
...@@ -12,12 +12,14 @@ ...@@ -12,12 +12,14 @@
#include "ck/library/tensor_operation_instance/gpu/normalization.hpp" #include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
using XDataType = ck::half_t; using XDataType = ck::half_t;
using GammaDataType = ck::half_t; using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t; using BetaDataType = ck::half_t;
using YDataType = ck::half_t; using YDataType = ck::half_t;
using ComputeDataType = float; using SaveMeanInvStdDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 2; constexpr int Rank = 2;
constexpr int NumReduceDim = 1; constexpr int NumReduceDim = 1;
...@@ -50,12 +52,16 @@ int main(int argc, char* argv[]) ...@@ -50,12 +52,16 @@ int main(int argc, char* argv[])
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N); SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N); SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size); SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
#ifdef SAVE_MEAN_INV_STD
SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * M);
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * M);
#endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType, using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
ComputeDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
PassThrough, PassThrough,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
...@@ -84,14 +90,21 @@ int main(int argc, char* argv[]) ...@@ -84,14 +90,21 @@ int main(int argc, char* argv[])
{0, 1}, // gammaStrides {0, 1}, // gammaStrides
{0, 1}, // betaStrides {0, 1}, // betaStrides
{Stride, 1}, // yStrides {Stride, 1}, // yStrides
{1}, // save_mean Strides
{1}, // save_inv_std Strides
{1}, // reduceDims {1}, // reduceDims
1e-4, 1e-4,
x_device_buf.GetDeviceBuffer(), x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(), gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(), beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(), y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr, nullptr,
nullptr, nullptr,
#endif
PassThrough{}); PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
...@@ -109,6 +122,10 @@ int main(int argc, char* argv[]) ...@@ -109,6 +122,10 @@ int main(int argc, char* argv[])
std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N + std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(YDataType) * M * N; sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
#ifdef SAVE_MEAN_INV_STD
num_byte += sizeof(SaveMeanInvStdDataType) * M * 2;
#endif
float gb_per_sec = num_byte / 1.E6 / ave_time; float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
...@@ -140,17 +157,24 @@ int main(int argc, char* argv[]) ...@@ -140,17 +157,24 @@ int main(int argc, char* argv[])
auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
{Stride, 1}, // xStrides {Stride, 1}, // xStrides
{1}, // gammaStrides {0, 1}, // gammaStrides
{1}, // betaStrides {0, 1}, // betaStrides
{Stride, 1}, // yStrides {Stride, 1}, // yStrides
{1}, // save_mean Strides
{1}, // save_inv_std Strides
{1}, // reduceDims {1}, // reduceDims
1e-4, 1e-4,
x_device_buf.GetDeviceBuffer(), x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(), gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(), beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(), y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr, nullptr,
nullptr, nullptr,
#endif
PassThrough{}); PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt deleted 100644 → 0
View file @ 755ace59
add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_operations)
client_example/10_grouped_convnd_bwd_data/CMakeLists.txt 0 → 100644
View file @ 9f8ab221
add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_conv3d_bwd_data grouped_conv3d_bwd_data.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 PRIVATE composable_kernel::device_operations)
client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp 0 → 100644
View file @ 9f8ab221
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 14;
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 14;
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
PassThrough,
PassThrough,
PassThrough>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Z * Y * X * C +
sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp 0 → 100644
View file @ 9f8ab221
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 14;
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 14;
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
PassThrough,
PassThrough,
PassThrough,
ck::bf8_t,
ck::f8_t>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Z * Y * X * C +
sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/16_convnd_fwd/CMakeLists.txt
View file @ 9f8ab221
add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp) if((DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp) add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp)
target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_operations) endif()
target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_operations)
if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
add_executable(client_conv3d_fwd_fp16_comp_fp8 conv3d_fwd_fp16_comp_fp8.cpp)
target_link_libraries(client_conv3d_fwd_fp16_comp_fp8 PRIVATE composable_kernel::device_operations)
endif()
if((DTYPES MATCHES "fp32") OR NOT DEFINED DTYPES)
add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp)
target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_operations)
endif()
client_example/16_convnd_fwd/common.hpp
View file @ 9f8ab221
...@@ -94,7 +94,8 @@ template <ck::index_t NumDimSpatial, ...@@ -94,7 +94,8 @@ template <ck::index_t NumDimSpatial,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
typename OutLayout, typename OutLayout,
ck::index_t NumNonSpatialDim = 3> ck::index_t NumNonSpatialDim = 3,
typename ComputeType = InDataType>
bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths, bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths, std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths) std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
...@@ -184,7 +185,8 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD ...@@ -184,7 +185,8 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough,
ComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp 0 → 100644
View file @ 9f8ab221
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 64;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
int main()
{
return run_grouped_conv_fwd<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
3,
ck::f8_t>(
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ 9f8ab221
...@@ -12,12 +12,14 @@ ...@@ -12,12 +12,14 @@
#include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp" #include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp"
using XDataType = ck::half_t; using XDataType = ck::half_t;
using GammaDataType = float; using GammaDataType = float;
using BetaDataType = float; using BetaDataType = float;
using YDataType = ck::half_t; using YDataType = ck::half_t;
using ComputeDataType = float; using SaveMeanInvStdDataType = float;
using Swish = ck::tensor_operation::element_wise::Swish; using Swish = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 5; constexpr int Rank = 5;
constexpr int NumReduceDim = 3; constexpr int NumReduceDim = 3;
...@@ -49,19 +51,24 @@ int main(int argc, char* argv[]) ...@@ -49,19 +51,24 @@ int main(int argc, char* argv[])
std::size_t xy_size = N * H * W * G * C; std::size_t xy_size = N * H * W * G * C;
std::size_t gamma_beta_size = G * C; std::size_t gamma_beta_size = G * C;
std::vector<ck::index_t> xy_strides = {H * W * G * C, W * G * C, G * C, C, 1}; std::vector<ck::index_t> xy_strides = {H * W * G * C, W * G * C, G * C, C, 1};
std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1}; std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
std::vector<ck::index_t> save_mean_inv_std_strides = {G, 1};
SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size); SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size); SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size); SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size); SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
#ifdef SAVE_MEAN_INV_STD
SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
#endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType, using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
ComputeDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
Swish, Swish,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
...@@ -75,19 +82,26 @@ int main(int argc, char* argv[]) ...@@ -75,19 +82,26 @@ int main(int argc, char* argv[])
const auto& generic_op_ptr = op_ptrs[0]; const auto& generic_op_ptr = op_ptrs[0];
auto generic_argument_ptr = auto generic_argument_ptr =
generic_op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths generic_op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides xy_strides, // xStrides
gamma_beta_strides, // gammaStrides gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides gamma_beta_strides, // betaStrides
xy_strides, // yStrides xy_strides, // yStrides
{1, 2, 4}, // reduceDims save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6, 1e-6,
x_device_buf.GetDeviceBuffer(), x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(), gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(), beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(), y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr, nullptr,
nullptr, nullptr,
#endif
Swish{}); Swish{});
if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get())) if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
...@@ -107,21 +121,29 @@ int main(int argc, char* argv[]) ...@@ -107,21 +121,29 @@ int main(int argc, char* argv[])
for(int i = 0; i < op_ptrs.size(); ++i) for(int i = 0; i < op_ptrs.size(); ++i)
{ {
auto& op_ptr = op_ptrs[i]; auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths auto argument_ptr =
xy_strides, // xStrides op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
gamma_beta_strides, // gammaStrides xy_strides, // xStrides
gamma_beta_strides, // betaStrides gamma_beta_strides, // gammaStrides
xy_strides, // yStrides gamma_beta_strides, // betaStrides
{1, 2, 4}, // reduceDims xy_strides, // yStrides
1e-6, save_mean_inv_std_strides, // save_mean Strides
x_device_buf.GetDeviceBuffer(), save_mean_inv_std_strides, // save_inv_std Strides
gamma_device_buf.GetDeviceBuffer(), {1, 2, 4}, // reduceDims
beta_device_buf.GetDeviceBuffer(), 1e-6,
y_device_buf.GetDeviceBuffer(), x_device_buf.GetDeviceBuffer(),
nullptr, gamma_device_buf.GetDeviceBuffer(),
nullptr, beta_device_buf.GetDeviceBuffer(),
Swish{}); y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
...@@ -139,6 +161,10 @@ int main(int argc, char* argv[]) ...@@ -139,6 +161,10 @@ int main(int argc, char* argv[])
sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size + sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size; sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
#ifdef SAVE_MEAN_INV_STD
num_byte += sizeof(SaveMeanInvStdDataType) * N * G * 2;
#endif
float gb_per_sec = num_byte / 1.E6 / ave_time; float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
...@@ -169,20 +195,28 @@ int main(int argc, char* argv[]) ...@@ -169,20 +195,28 @@ int main(int argc, char* argv[])
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString() std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl; << std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths auto argument_ptr =
xy_strides, // xStrides op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
gamma_beta_strides, // gammaStrides xy_strides, // xStrides
gamma_beta_strides, // betaStrides gamma_beta_strides, // gammaStrides
xy_strides, // yStrides gamma_beta_strides, // betaStrides
{1, 2, 4}, // reduceDims xy_strides, // yStrides
1e-6, save_mean_inv_std_strides, // save_mean Strides
x_device_buf.GetDeviceBuffer(), save_mean_inv_std_strides, // save_inv_std Strides
gamma_device_buf.GetDeviceBuffer(), {1, 2, 4}, // reduceDims
beta_device_buf.GetDeviceBuffer(), 1e-6,
y_device_buf.GetDeviceBuffer(), x_device_buf.GetDeviceBuffer(),
nullptr, gamma_device_buf.GetDeviceBuffer(),
nullptr, beta_device_buf.GetDeviceBuffer(),
Swish{}); y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
View file @ 9f8ab221
...@@ -60,14 +60,13 @@ int main() ...@@ -60,14 +60,13 @@ int main()
int sum_of_m = 0; int sum_of_m = 0;
Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148}; const int group_count = 16;
int group_count = Ms.size();
for(int i = 0; i < group_count; ++i) for(int i = 0; i < group_count; ++i)
{ {
Ns.push_back(768); Ms.push_back(256 + 256 * i);
Ks.push_back(4608); Ns.push_back(128 + 128 * i);
Ks.push_back(128 + 64 * i);
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]); StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]); StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
......
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
View file @ 9f8ab221
...@@ -57,15 +57,13 @@ int main() ...@@ -57,15 +57,13 @@ int main()
int sum_of_m = 0; int sum_of_m = 0;
// Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148}; const int group_count = 16;
Ms = {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0};
int group_count = Ms.size();
for(int i = 0; i < group_count; ++i) for(int i = 0; i < group_count; ++i)
{ {
Ns.push_back(768); Ms.push_back(256 + 256 * i);
Ks.push_back(4608); Ns.push_back(128 + 128 * i);
Ks.push_back(128 + 64 * i);
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]); StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]); StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
......
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
View file @ 9f8ab221
...@@ -58,14 +58,13 @@ int main() ...@@ -58,14 +58,13 @@ int main()
int sum_of_m = 0; int sum_of_m = 0;
Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148}; const int group_count = 16;
int group_count = Ms.size();
for(int i = 0; i < group_count; ++i) for(int i = 0; i < group_count; ++i)
{ {
Ns.push_back(768); Ms.push_back(256 + 256 * i);
Ks.push_back(4608); Ns.push_back(128 + 128 * i);
Ks.push_back(128 + 64 * i);
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]); StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]); StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
......
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
View file @ 9f8ab221
...@@ -58,14 +58,13 @@ int main() ...@@ -58,14 +58,13 @@ int main()
int sum_of_m = 0; int sum_of_m = 0;
Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148}; const int group_count = 16;
int group_count = Ms.size();
for(int i = 0; i < group_count; ++i) for(int i = 0; i < group_count; ++i)
{ {
Ns.push_back(768); Ms.push_back(256 + 256 * i);
Ks.push_back(4608); Ns.push_back(128 + 128 * i);
Ks.push_back(128 + 64 * i);
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]); StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]); StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
......
client_example/20_image_to_column/CMakeLists.txt → client_example/22_im2col_col2im/CMakeLists.txt
View file @ 9f8ab221
add_executable(client_image_to_column image_to_column.cpp) add_executable(client_image_to_column image_to_column.cpp)
target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations) target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)
add_executable(client_column_to_image column_to_image.cpp)
target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_operations)
client_example/22_im2col_col2im/column_to_image.cpp 0 → 100644
View file @ 9f8ab221
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp"
#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using ImageLayout = ck::tensor_layout::convolution::GNHWC;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 32; // batch size
static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 28; // input H
static constexpr ck::index_t Wi = 28; // input W
static constexpr ck::index_t Ho = 28; // output H
static constexpr ck::index_t Wo = 28; // output W
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, 2> in_spatial_lengths{Hi, Wi};
std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
// We have NHWGC in memory space (G is dummy)
// However, CK's API only accept length and stride with order of GNCHW
// Hence, we need to adjust the order of stride
std::array<ck::index_t, 5> image_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
std::array<ck::index_t, 2> gemm_strides{Y * X * C, 1};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Ho * Wo * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Hi * Wi * G * C);
using namespace ck::conv_tensor_rearrange_op;
using DeviceOp = ck::tensor_operation::device::DeviceConvTensorRearrange<NumDimSpatial,
ImageLayout,
InDataType,
OutDataType,
ColumnToImage>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
out.GetDeviceBuffer(),
N,
C,
in_spatial_lengths,
out_spatial_lengths,
wei_spatial_lengths,
image_strides,
gemm_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
sizeof(OutDataType) * N * Ho * Wo * Y * X * C;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(avg_time < best_avg_time)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
out.GetDeviceBuffer(),
N,
C,
in_spatial_lengths,
out_spatial_lengths,
wei_spatial_lengths,
image_strides,
gemm_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
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