Commit f6ceef78 authored by ThomasNing's avatar ThomasNing
Browse files

merge with the develop branch

parents 536c5458 25935b57
......@@ -62,8 +62,14 @@ if (DTYPES)
endif()
message("DTYPES macro set to ${DTYPES}")
else()
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")
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8)
set(CK_ENABLE_INT8 "ON")
set(CK_ENABLE_FP16 "ON")
set(CK_ENABLE_FP32 "ON")
set(CK_ENABLE_FP64 "ON")
set(CK_ENABLE_BF16 "ON")
set(CK_ENABLE_FP8 "ON")
set(CK_ENABLE_BF8 "ON")
endif()
#for f8/bf8_t type
......@@ -106,21 +112,33 @@ list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/ll
message("GPU_TARGETS= ${GPU_TARGETS}")
find_package(hip)
# No assumption that HIP kernels are launched with uniform block size for backward compatibility
# SWDEV-413293 and https://reviews.llvm.org/D155213
math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}")
message("hip_version_flat=${hip_VERSION_FLAT}")
message("checking which targets are supported")
#This is the list of targets to be used in case GPU_TARGETS is not set on command line
#These targets will be filtered and only supported ones will be used
#Setting GPU_TARGETS on command line will override this list
if(NOT PROFILER_ONLY)
if(NOT ENABLE_ASAN_PACKAGING)
#build CK for all supported targets
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
else()
#build CK only for xnack-supported targets
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
endif()
if(NOT ENABLE_ASAN_PACKAGING)
#build CK for all supported targets
if(NOT WIN32 AND ${hip_VERSION_FLAT} LESS 600300000)
# WORKAROUND: compiler does not yet fully support gfx12 targets, need to fix version above
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
else()
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
endif()
else()
#build CK only for xnack-supported targets
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
endif()
else()
add_definitions(-DPROFILER_ONLY)
set(GPU_TARGETS "" CACHE STRING "" FORCE)
......@@ -169,18 +187,30 @@ endif()
# CK config file to record supported datatypes, etc.
configure_file(include/ck/config.h.in ${CMAKE_CURRENT_BINARY_DIR}/include/ck/config.h)
find_package(hip)
# No assumption that HIP kernels are launched with uniform block size for backward compatibility
# SWDEV-413293 and https://reviews.llvm.org/D155213
math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}")
message("hip_version_flat=${hip_VERSION_FLAT}")
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 500723302)
message("Adding the fno-offload-uniform-block compiler flag")
add_compile_options(-fno-offload-uniform-block)
check_cxx_compiler_flag("-fno-offload-uniform-block" HAS_NO_OFFLOAD_UNIFORM_BLOCK)
if(HAS_NO_OFFLOAD_UNIFORM_BLOCK)
message("Adding the fno-offload-uniform-block compiler flag")
add_compile_options(-fno-offload-uniform-block)
endif()
endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
message("Adding the enable-post-misched=0 compiler flag")
add_compile_options(-mllvm -enable-post-misched=0)
check_cxx_compiler_flag("-mllvm -enable-post-misched=0" HAS_ENABLE_POST_MISCHED)
if(HAS_ENABLE_POST_MISCHED)
message("Adding the enable-post-misched=0 compiler flag")
add_compile_options("SHELL: -mllvm -enable-post-misched=0")
endif()
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)
message("Adding the amdgpu-coerce-illegal-types=1")
add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding -amdgpu-early-inline-all=true and -amdgpu-function-calls=false")
add_compile_options("SHELL: -mllvm -amdgpu-early-inline-all=true")
add_compile_options("SHELL: -mllvm -amdgpu-function-calls=false")
endif()
#
# Seperate linking jobs from compiling
......@@ -235,6 +265,16 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)
message("CMAKE_CXX_COMPILER: ${CMAKE_CXX_COMPILER}")
# https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_macros.html
# _GLIBCXX_ASSERTIONS
# Undefined by default. When defined, enables extra error checking in the form of
# precondition assertions, such as bounds checking in strings and null pointer
# checks when dereferencing smart pointers
option(USE_GLIBCXX_ASSERTIONS "Turn on additional c++ library checks." OFF)
if(USE_GLIBCXX_ASSERTIONS)
add_compile_options(-Wp,-D_GLIBCXX_ASSERTIONS)
endif()
## HIP
set(CMAKE_HIP_PLATFORM amd)
set(CMAKE_HIP_COMPILER ${CMAKE_CXX_COMPILER})
......@@ -513,9 +553,7 @@ if(NOT DEFINED INSTANCES_ONLY)
PACKAGE_NAME examples
)
add_subdirectory(example)
if(BUILD_TESTING)
add_subdirectory(test)
endif()
add_subdirectory(test)
rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel
......@@ -532,6 +570,10 @@ if(NOT DEFINED INSTANCES_ONLY)
endif()
endif()
if(NOT DEFINED PROFILER_ONLY AND (GPU_TARGETS MATCHES "gfx9" OR DEFINED INSTANCES_ONLY))
add_subdirectory(codegen)
endif()
#Create an interface target for the include only files and call it "composablekernels"
include(CMakePackageConfigHelpers)
......
FROM ubuntu:20.04
ARG DEBIAN_FRONTEND=noninteractive
ARG ROCMVERSION=6.1
ARG ROCMVERSION=6.2
ARG compiler_version=""
ARG compiler_commit=""
ARG CK_SCCACHE=""
......@@ -17,17 +17,12 @@ RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN if [ "$ROCMVERSION" != "6.2" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/6.1/ubuntu/focal/amdgpu-install_6.1.60100-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.1.60100-1_all.deb && \
RUN if [ "$ROCMVERSION" != "6.3" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.2.60200-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.2.60200-1_all.deb && \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "6.2" ] && [ "$compiler_version" = "rc3" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.2-20.04-1_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_6.2-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.2 rel-45 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2003709; \
fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
......@@ -64,6 +59,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
python3-dev \
python3-pip \
redis \
rocm-llvm-dev \
sshpass \
stunnel \
software-properties-common \
......
......@@ -38,7 +38,7 @@ def getDockerImageName(){
img = "${params.USE_CUSTOM_DOCKER}"
}
else{
if (params.ROCMVERSION != "6.2"){
if (params.ROCMVERSION != "6.3"){
if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
......@@ -86,17 +86,7 @@ def check_host() {
def build_compiler(){
def compiler
if (params.BUILD_COMPILER == "hipcc"){
compiler = '/opt/rocm/bin/hipcc'
}
else{
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
compiler = "/llvm-project/build/bin/clang++"
}
else{
compiler = "/opt/rocm/llvm/bin/clang++"
}
}
compiler = "${params.BUILD_COMPILER}"
return compiler
}
......@@ -214,6 +204,9 @@ def cmake_build(Map conf=[:]){
cd build
"""
def invocation_tag=""
if (setup_args.contains("gfx12")){
invocation_tag="gfx12"
}
if (setup_args.contains("gfx11")){
invocation_tag="gfx11"
}
......@@ -295,6 +288,19 @@ def cmake_build(Map conf=[:]){
if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) {
archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true
}
if (params.RUN_CK_TILE_TESTS){
try{
archiveArtifacts "perf_fmha_fwd_*.log"
archiveArtifacts "perf_fmha_bwd_*.log"
stash name: "perf_fmha_fwd_gfx942.log"
stash name: "perf_fmha_bwd_gfx942.log"
stash name: "perf_fmha_fwd_gfx90a.log"
stash name: "perf_fmha_bwd_gfx90a.log"
}
catch(Exception err){
echo "could not locate the requested artifacts: ${err.getMessage()}. will skip the stashing."
}
}
}
def buildHipClangJob(Map conf=[:]){
......@@ -307,7 +313,7 @@ def buildHipClangJob(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
......@@ -366,7 +372,7 @@ def runCKProfiler(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
......@@ -420,8 +426,9 @@ def runCKProfiler(Map conf=[:]){
archiveArtifacts "perf_resnet50_N4.log"
archiveArtifacts "perf_batched_gemm.log"
archiveArtifacts "perf_grouped_gemm.log"
archiveArtifacts "perf_conv_fwd.log"
archiveArtifacts "perf_conv_bwd_data.log"
archiveArtifacts "perf_grouped_conv_fwd.log"
archiveArtifacts "perf_grouped_conv_bwd_data.log"
archiveArtifacts "perf_grouped_conv_bwd_weight.log"
archiveArtifacts "perf_gemm_bilinear.log"
archiveArtifacts "perf_reduction.log"
archiveArtifacts "perf_splitK_gemm.log"
......@@ -433,8 +440,9 @@ def runCKProfiler(Map conf=[:]){
stash name: "perf_resnet50_N4.log"
stash name: "perf_batched_gemm.log"
stash name: "perf_grouped_gemm.log"
stash name: "perf_conv_fwd.log"
stash name: "perf_conv_bwd_data.log"
stash name: "perf_grouped_conv_fwd.log"
stash name: "perf_grouped_conv_bwd_data.log"
stash name: "perf_grouped_conv_bwd_weight.log"
stash name: "perf_gemm_bilinear.log"
stash name: "perf_reduction.log"
stash name: "perf_splitK_gemm.log"
......@@ -487,7 +495,7 @@ def Build_CK(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
......@@ -528,7 +536,7 @@ def Build_CK(Map conf=[:]){
//check whether to run performance tests on this node
def do_perf_tests = 0
sh 'rocminfo | tee rocminfo.log'
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx1201" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
do_perf_tests = 1
echo "Stash profiler and run performance tests"
}
......@@ -600,7 +608,7 @@ def process_results(Map conf=[:]){
def prefixpath = "/opt/rocm"
// Jenkins is complaining about the render group
def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
def dockerOpts="--rm --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
......@@ -622,6 +630,17 @@ def process_results(Map conf=[:]){
timeout(time: 1, unit: 'HOURS'){
try{
dir("script"){
if (params.RUN_CK_TILE_TESTS){
try{
unstash "perf_fmha_fwd_gfx942.log"
unstash "perf_fmha_bwd_gfx942.log"
unstash "perf_fmha_fwd_gfx90a.log"
unstash "perf_fmha_bwd_gfx90a.log"
}
catch(Exception err){
echo "could not locate the FMHA performance logs: ${err.getMessage()}."
}
}
if (params.RUN_FULL_QA){
// unstash perf files to master
unstash "ckprofiler_0.2.0_amd64.deb"
......@@ -631,8 +650,9 @@ def process_results(Map conf=[:]){
unstash "perf_resnet50_N4.log"
unstash "perf_batched_gemm.log"
unstash "perf_grouped_gemm.log"
unstash "perf_conv_fwd.log"
unstash "perf_conv_bwd_data.log"
unstash "perf_grouped_conv_fwd.log"
unstash "perf_grouped_conv_bwd_data.log"
unstash "perf_grouped_conv_bwd_weight.log"
unstash "perf_gemm_bilinear.log"
unstash "perf_reduction.log"
unstash "perf_splitK_gemm.log"
......@@ -662,10 +682,10 @@ def process_results(Map conf=[:]){
}
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.1; RUN_CK_TILE_TESTS=true
0 21 * * * % ROCMVERSION=6.1;hipTensor_test=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;COMPILER_COMMIT=;USE_SCCACHE=false
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;COMPILER_COMMIT=;USE_SCCACHE=false
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2; RUN_CK_TILE_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=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
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
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
pipeline {
......@@ -687,8 +707,8 @@ pipeline {
description: 'If you want to use a custom docker image, please specify it here (default: leave blank).')
string(
name: 'ROCMVERSION',
defaultValue: '6.1',
description: 'Specify which ROCM version to use: 6.1 (default).')
defaultValue: '6.2',
description: 'Specify which ROCM version to use: 6.2 (default).')
string(
name: 'COMPILER_VERSION',
defaultValue: '',
......@@ -699,8 +719,8 @@ pipeline {
description: 'Specify which commit of compiler branch to use: leave blank to use the latest commit (default), or use some specific commit of llvm-project branch.')
string(
name: 'BUILD_COMPILER',
defaultValue: 'clang',
description: 'Specify whether to build CK with hipcc or with clang (default).')
defaultValue: '/opt/rocm/llvm/bin/clang++',
description: 'Build CK with /opt/rocm/bin/hipcc, /llvm-project/build/bin/clang++, or with /opt/rocm/llvm/bin/clang++ (default).')
booleanParam(
name: "RUN_FULL_QA",
defaultValue: false,
......@@ -730,9 +750,9 @@ pipeline {
defaultValue: true,
description: "Run the performance tests (default: ON)")
booleanParam(
name: "RUN_CODEGEN_TESTS",
defaultValue: true,
description: "Run the codegen tests (default: ON)")
name: "RUN_GROUPED_CONV_LARGE_CASES_TESTS",
defaultValue: false,
description: "Run the grouped conv large cases tests (default: OFF)")
booleanParam(
name: "RUN_CK_TILE_TESTS",
defaultValue: false,
......@@ -741,6 +761,11 @@ pipeline {
name: "BUILD_INSTANCES_ONLY",
defaultValue: false,
description: "Test building instances for various architectures simultaneously (default: OFF)")
booleanParam(
name: "BUILD_GFX12",
defaultValue: false,
description: "Build CK and run tests on gfx12 (default: OFF)")
}
environment{
dbuser = "${dbuser}"
......@@ -819,25 +844,22 @@ pipeline {
}
}
}
stage("Run Codegen Tests")
stage("Run Grouped Conv Large Case Tests")
{
parallel
{
stage("Run Codegen Tests on gfx90a")
stage("Run Grouped Conv Large Case Tests on gfx90a")
{
when {
beforeAgent true
expression { params.RUN_CODEGEN_TESTS.toBoolean() }
expression { params.RUN_GROUPED_CONV_LARGE_CASES_TESTS.toBoolean() }
}
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "NO_CK_BUILD"
execute_args = """ cd ../codegen && rm -rf build && mkdir build && cd build && \
cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ \
-D CMAKE_BUILD_TYPE=Release \
-D GPU_TARGETS="gfx90a" \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j check"""
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx90a && \
make -j64 test_grouped_convnd_fwd_large_cases_xdl && \
./bin/test_grouped_convnd_fwd_large_cases_xdl"""
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
......@@ -862,8 +884,7 @@ pipeline {
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx90a && \
make -j64 tile_example_fmha_fwd tile_example_fmha_bwd && \
cd ../ &&
example/ck_tile/01_fmha/script/smoke_test_fwd.sh && \
example/ck_tile/01_fmha/script/smoke_test_bwd.sh"""
example/ck_tile/01_fmha/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx90a """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
......@@ -882,8 +903,7 @@ pipeline {
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx942 && \
make -j64 tile_example_fmha_fwd tile_example_fmha_bwd && \
cd ../ &&
example/ck_tile/01_fmha/script/smoke_test_fwd.sh && \
example/ck_tile/01_fmha/script/smoke_test_bwd.sh"""
example/ck_tile/01_fmha/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx942 """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
......@@ -1018,6 +1038,26 @@ pipeline {
cleanWs()
}
}
stage("Build CK and run Tests on gfx1201")
{
when {
beforeAgent true
expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
}
agent{ label rocmnode("gfx1201") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1201" -DDL_KERNELS=ON -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx1201" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
cleanWs()
}
}
}
}
......
......@@ -5,17 +5,17 @@ if(GPU_TARGETS MATCHES "gfx9")
add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_conv_operations)
if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_fwd_fp8 grouped_conv3d_fwd_fp8.cpp)
target_link_libraries(client_grouped_conv3d_fwd_fp8 PRIVATE composable_kernel::device_conv_operations)
endif()
if((DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_fwd_bf8 grouped_conv3d_fwd_bf8.cpp)
target_link_libraries(client_grouped_conv3d_fwd_bf8 PRIVATE composable_kernel::device_conv_operations)
endif()
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_fwd_fp8_bf8 grouped_conv3d_fwd_fp8_bf8.cpp)
target_link_libraries(client_grouped_conv3d_fwd_fp8_bf8 PRIVATE composable_kernel::device_conv_operations)
......
......@@ -4,5 +4,7 @@ target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::
add_executable(client_grouped_conv3d_bwd_data grouped_conv3d_bwd_data.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data PRIVATE composable_kernel::device_conv_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_conv_operations)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
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_conv_operations)
endif()
\ No newline at end of file
......@@ -2,10 +2,13 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
endif()
\ No newline at end of file
......@@ -4,7 +4,7 @@ if((DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
endif()
if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
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_conv_operations)
endif()
......
if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES))
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_gemm_operations)
endif()
if(GPU_TARGETS MATCHES "gfx9")
# Fwd scaleadd scaleadd relu
add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32
add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32
grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 PRIVATE composable_kernel::device_conv_operations)
......@@ -36,7 +36,7 @@ add_executable(client_grouped_convnd_fwd_bilinear_residual_fp16
grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp)
target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
# Fwd convinvscale
add_executable(client_conv3d_fwd_convinvscale_fp8
add_executable(client_conv3d_fwd_convinvscale_fp8
grouped_convnd_fwd_convinvscale/conv3d_fwd_convinvscale_fp8.cpp)
target_link_libraries(client_conv3d_fwd_convinvscale_fp8 PRIVATE composable_kernel::device_conv_operations)
# Fwd convscale + Bias
......@@ -47,6 +47,22 @@ target_link_libraries(client_conv3d_fwd_convscale_add_fp8 PRIVATE composable_ker
add_executable(client_conv3d_fwd_convscale_relu_fp8
grouped_convnd_fwd_convscale_relu/conv3d_fwd_convscale_relu_fp8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_relu_fp8 PRIVATE composable_kernel::device_conv_operations)
# Fwd convscale + ReLU + AMAX
add_executable(client_conv3d_fwd_convscale_relu_amax_fp8
grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_relu_amax_fp8.cpp)
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)
# Fwd convscale + AMAX
add_executable(client_conv3d_fwd_convscale_amax_fp8
grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_amax_fp8.cpp)
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)
# Fwd convscale
add_executable(client_conv3d_fwd_convscale_fp8
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8.cpp)
......@@ -56,11 +72,11 @@ add_executable(client_conv3d_fwd_convscale_bf8
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_bf8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_bf8 PRIVATE composable_kernel::device_conv_operations)
add_executable(client_conv3d_fwd_convscale_fp8_bf8
add_executable(client_conv3d_fwd_convscale_fp8_bf8
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8_bf8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_fp8_bf8 PRIVATE composable_kernel::device_conv_operations)
add_executable(client_conv3d_fwd_convscale_bf8_fp8
add_executable(client_conv3d_fwd_convscale_bf8_fp8
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_bf8_fp8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
# Bwd data bilinear
......
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <string>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/utility/type.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_relu.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/permute_scale.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
#include "ck/library/utility/host_tensor.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ConvScaleRelu = ck::tensor_operation::element_wise::ScaleScaleRelu;
using ConvScale = ck::tensor_operation::element_wise::ScaleScalePass;
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_;
};
template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
const std::size_t& ds_size)
{
// 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product> +
// + ds_size * <output tensor size> =>
// => <output tensor size> * ( 2 * C * <filter spatial lengths product> + ds_size) =>
// => G * N * K * <output spatial lengths product> * (2 * C * <filter spatial lengths product> +
// ds_size)
ck::index_t G = weights_lengths[0];
ck::index_t N = output_lengths[1];
ck::index_t K = weights_lengths[1];
ck::index_t C = weights_lengths[2];
return G * N * K *
std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
std::end(output_lengths),
static_cast<std::size_t>(1),
std::multiplies<>()) *
(ds_size + static_cast<std::size_t>(2) * C *
std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
std::end(weights_lengths),
static_cast<std::size_t>(1),
std::multiplies<>()));
}
template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t GetTensorSize(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths)
{
return std::accumulate(std::begin(lengths),
std::end(lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>());
}
template <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
{
// sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
return sizeof(InDataType) * GetTensorSize<NumDimSpatial>(input_lengths);
}
template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
{
// sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
return sizeof(WeiDataType) * GetTensorSize<NumDimSpatial>(weights_lengths);
}
template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
{
// sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
return sizeof(OutDataType) * GetTensorSize<NumDimSpatial>(output_lengths);
}
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename ConvElementOp,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim = 3,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool ConvolutionScale(SimpleDeviceMem& in,
SimpleDeviceMem& wei,
SimpleDeviceMem& out,
ConvElementOp elementwise_op,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_strides);
template <typename InDataType,
typename OutDataType,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim = 3>
bool TensorScaleConvert(SimpleDeviceMem& in,
SimpleDeviceMem& out,
float scale_out,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides);
template <typename InDataType,
typename OutDataType,
ck::ReduceTensorOp ReduceOpId,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim = 3>
bool TensorFullReduction(SimpleDeviceMem& tensor,
SimpleDeviceMem& out_amax,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides);
template <ck::index_t NumDimSpatial,
typename InDataType,
typename WeiDataType,
typename ConvOutDataType,
typename OutDataType,
typename ConvElementOp,
ck::ReduceTensorOp ReduceOp,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumNonSpatialDim = 3,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_fwd_convscale_reduce(
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> out_lengths)
{
namespace ctc = ck::tensor_layout::convolution;
static_assert(NumDimSpatial == 3 && ck::is_same_v<InLayout, ctc::NDHWGC> &&
ck::is_same_v<WeiLayout, ctc::GKZYXC> &&
ck::is_same_v<OutLayout, ctc::NDHWGK>,
"Unsupported configuration");
const ck::index_t G = in_lengths[4];
const ck::index_t N = in_lengths[0];
const ck::index_t K = wei_lengths[1];
const ck::index_t C = in_lengths[5];
const ck::index_t Z = wei_lengths[2];
const ck::index_t Y = wei_lengths[3];
const ck::index_t X = wei_lengths[4];
const ck::index_t Di = in_lengths[1];
const ck::index_t Hi = in_lengths[2];
const ck::index_t Wi = in_lengths[3];
const ck::index_t Do = out_lengths[1];
const ck::index_t Ho = out_lengths[2];
const ck::index_t Wo = out_lengths[3];
const std::size_t in_mem_size = sizeof(InDataType) * N * Di * Hi * Wi * G * C;
const std::size_t wei_mem_size = sizeof(WeiDataType) * G * K * Z * Y * X * C;
const std::size_t conv_out_mem_size = sizeof(ConvOutDataType) * N * Do * Ho * Wo * G * K;
const std::size_t out_mem_size = sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
SimpleDeviceMem in(in_mem_size);
SimpleDeviceMem wei(wei_mem_size);
SimpleDeviceMem conv_out(conv_out_mem_size);
SimpleDeviceMem out(out_mem_size);
float scale_in = float(std::rand()) / float(RAND_MAX);
float scale_wei = float(std::rand()) / float(RAND_MAX);
float scale_out = float(std::rand()) / float(RAND_MAX);
// We have NDHWGC/GKZYXC/NDHWGK (x, weight, y) in memory space.
// However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
// Hence, we need to adjust the order of strides.
const std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
const std::array<ck::index_t, NumDimSpatial + 3> input_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
const std::array<ck::index_t, NumDimSpatial + 3> weights_lengths{G, K, C, Z, Y, X};
const std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
const std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
const std::array<ck::index_t, NumDimSpatial + 3> output_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
/*
* FP8 Convolution with Scaling
*/
std::cout << "\n\nConvolution with scale Benchmarking:" << std::endl;
auto elementwise_op = ConvElementOp{ck::tensor_operation::element_wise::Scale{scale_in},
ck::tensor_operation::element_wise::Scale{scale_wei},
{}};
auto conv_ok = ConvolutionScale<InDataType,
WeiDataType,
ConvOutDataType,
ConvElementOp,
InLayout,
WeiLayout,
OutLayout,
NumDimSpatial>(in,
wei,
conv_out,
elementwise_op,
input_lengths,
input_strides,
weights_lengths,
weights_strides,
output_lengths,
output_strides);
if(!conv_ok)
return false;
/*
* Scale with output weight and convert to FP8
*/
std::cout << "\n\nElement-wise scale + convert Benchmarking:" << std::endl;
auto elem_wise_ok = TensorScaleConvert<ConvOutDataType, OutDataType, NumDimSpatial>(
conv_out, out, scale_out, output_lengths, output_strides);
if(!elem_wise_ok)
return false;
/*
* Compute AMAX
*/
std::cout << "\n\nAMAX Benchmarking:" << std::endl;
SimpleDeviceMem amax_device(sizeof(ConvOutDataType));
auto reduction_ok =
TensorFullReduction<ConvOutDataType,
ConvOutDataType,
ck::ReduceTensorOp::AMAX,
NumDimSpatial>(conv_out, amax_device, output_lengths, output_strides);
if(!reduction_ok)
return false;
return true;
}
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename ConvElementOp,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim,
typename AComputeType,
typename BComputeType>
bool ConvolutionScale(SimpleDeviceMem& in,
SimpleDeviceMem& wei,
SimpleDeviceMem& out,
ConvElementOp elementwise_op,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_strides)
{
const std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
const std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
const std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
const std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
const auto in_mem_size = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
const auto wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
const auto out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
std::size_t ds_size = 2; // 2 element-wise scale multipliers
if constexpr(ck::is_same_v<ConvElementOp, ConvScaleRelu>)
{
ds_size += 1; // +1 element-wise relu
}
std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
std::size_t num_bytes =
in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + out_mem_size;
using ConvDeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
ConvElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto conv_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
ConvDeviceOp>::GetInstances();
std::cout << "found " << conv_ptrs.size() << " instances" << std::endl;
std::string conv_best_op_name;
int conv_best_op_id = -1;
float conv_best_avg_time = std::numeric_limits<float>::max();
float conv_best_gb_per_sec = 0;
float conv_best_tflops = 0;
// profile device operation instances
std::cout << "Run all convolution instances and do timing" << std::endl;
for(int i = 0; i < conv_ptrs.size(); ++i)
{
auto& op_ptr = conv_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
std::array<const void*, 0>{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
out_lengths,
out_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
elementwise_op);
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});
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 > conv_best_tflops)
{
conv_best_op_id = i;
conv_best_op_name = op_name;
conv_best_avg_time = avg_time;
conv_best_gb_per_sec = gb_per_sec;
conv_best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(conv_best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return false;
}
std::cout << "Best Perf: " << std::setw(10) << conv_best_avg_time << " ms, " << conv_best_tflops
<< " TFlops, " << conv_best_gb_per_sec << " GB/s, " << conv_best_op_name << std::endl;
// run the best instance
{
auto& op_ptr = conv_ptrs[conv_best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
std::array<const void*, 0>{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
out_lengths,
out_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
elementwise_op);
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;
}
return true;
}
template <typename InDataType,
typename OutDataType,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim>
bool TensorScaleConvert(SimpleDeviceMem& in,
SimpleDeviceMem& out,
float scale_out,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides)
{
const auto tensor_size = GetTensorSize<NumDimSpatial>(lengths);
const std::size_t in_mem_size = sizeof(InDataType) * tensor_size;
const std::size_t out_mem_size = sizeof(OutDataType) * tensor_size;
std::size_t flop = 2 * tensor_size; // element-wise scale + convert
std::size_t bytes =
in_mem_size + sizeof(float) + out_mem_size; // read from in, scale, write to out
using DeviceScaleConvert =
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<InDataType>,
ck::Tuple<OutDataType>,
ck::tensor_operation::element_wise::Scale,
NumDimSpatial + NumNonSpatialDim>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceScaleConvert>::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 DeviceScaleConvert instances and do timing" << std::endl;
auto scale_convert = ck::tensor_operation::element_wise::Scale{scale_out};
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
{strides},
{strides},
{in.GetDeviceBuffer()},
{out.GetDeviceBuffer()},
scale_convert);
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});
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = 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 found." << std::endl;
return false;
}
else
{
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(lengths,
{strides},
{strides},
{in.GetDeviceBuffer()},
{out.GetDeviceBuffer()},
scale_convert);
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;
}
return true;
}
template <typename InDataType,
typename OutDataType,
ck::ReduceTensorOp ReduceOpId,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim>
bool TensorFullReduction(SimpleDeviceMem& tensor,
SimpleDeviceMem& out_amax,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides)
{
const auto spatial_dim_size = std::accumulate(std::next(std::begin(lengths), NumNonSpatialDim),
std::end(lengths),
static_cast<std::size_t>(1),
std::multiplies<>());
const auto tensor_size = GetTensorSize<NumDimSpatial>(lengths);
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
// Get the reduction operation
using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation =
typename ck::reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename ck::reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
InElementwiseOperation in_elementwise_op;
AccElementwiseOperation acc_elementwise_op;
std::tie(in_elementwise_op, acc_elementwise_op) =
ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(tensor_size));
std::array<ck::index_t, 1> reduce_out_lengths{1};
std::array<ck::index_t, 1> reduce_out_strides{1};
SimpleDeviceMem partial_reduce_tensor(sizeof(OutDataType) * spatial_dim_size);
std::array<ck::index_t, NumDimSpatial> reduce_part_lengths;
std::copy(std::next(std::begin(lengths), NumNonSpatialDim),
std::end(lengths),
std::begin(reduce_part_lengths));
std::array<ck::index_t, NumDimSpatial> reduce_part_strides;
copy(HostTensorDescriptor(reduce_part_lengths).GetStrides(), reduce_part_strides);
{
std::cout << "\nReduction of nonspatial dimensions:" << std::endl;
using DeviceOp =
ck::tensor_operation::device::DeviceReduce<InDataType,
OutDataType,
OutDataType,
NumDimSpatial + NumNonSpatialDim,
NumNonSpatialDim,
ReduceOperation,
InElementwiseOperation,
PassThrough,
true, // PropagateNan
false>; // OutputIndex
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_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
std::array<int, NumNonSpatialDim> reduce_dims;
std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NumNonSpatialDim-1
ck::index_t num_in_elements = tensor_size;
ck::index_t num_out_elements = spatial_dim_size;
// profile device operation instances
std::cout << "Run partial reduction 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(lengths,
strides,
reduce_part_lengths,
reduce_part_strides,
reduce_dims,
1.0,
0.0,
tensor.GetDeviceBuffer(),
nullptr,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
in_elementwise_op,
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes =
num_in_elements * sizeof(InDataType) + num_out_elements * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(ave_time < best_ave_time)
{
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance found." << std::endl;
return false;
}
else
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best instance
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(lengths,
strides,
reduce_part_lengths,
reduce_part_strides,
reduce_dims,
1.0,
0.0,
tensor.GetDeviceBuffer(),
nullptr,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
in_elementwise_op,
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;
}
}
{
std::cout << "\nReduction of spatial dimensions:" << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceReduce<OutDataType,
OutDataType,
OutDataType,
NumDimSpatial,
NumDimSpatial,
ReduceOperation,
PassThrough,
AccElementwiseOperation,
true, // PropagateNan
false>; // OutputIndex
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_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
std::array<int, NumDimSpatial> reduce_dims;
std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NumDimSpatial-1
ck::index_t num_in_elements = spatial_dim_size;
ck::index_t num_out_elements = 1;
// profile device operation instances
std::cout << "Run final reduction 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(reduce_part_lengths,
reduce_part_strides,
reduce_out_lengths,
reduce_out_strides,
reduce_dims,
1.0,
0.0,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
out_amax.GetDeviceBuffer(),
nullptr,
PassThrough{},
acc_elementwise_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes =
num_in_elements * sizeof(OutDataType) + num_out_elements * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(ave_time < best_ave_time)
{
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance found." << std::endl;
return false;
}
else
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best instance
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(reduce_part_lengths,
reduce_part_strides,
reduce_out_lengths,
reduce_out_strides,
reduce_dims,
1.0,
0.0,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
out_amax.GetDeviceBuffer(),
nullptr,
PassThrough{},
acc_elementwise_op);
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;
}
}
return true;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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::f8_t;
using WeiDataType = ck::f8_t;
using CShuffleDataType = float;
using ConvOutDataType = float; // data type of convolution result
using OutDataType = ck::f8_t; // data type of final result
using AComputeDataType = ck::f8_t;
using BComputeDataType = ck::f8_t;
using ConvElementOp = ConvScale;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
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_convscale_reduce<NumDimSpatial,
InDataType,
WeiDataType,
ConvOutDataType,
OutDataType,
ConvElementOp,
ReduceOpId,
InLayout,
WeiLayout,
OutLayout,
3,
AComputeDataType,
BComputeDataType>(
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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::f8_t;
using WeiDataType = ck::f8_t;
using CShuffleDataType = float;
using ConvOutDataType = float; // data type of convolution result
using OutDataType = ck::f8_t; // data type of final result
using AComputeDataType = ck::f8_t;
using BComputeDataType = ck::f8_t;
using ConvElementOp = ConvScaleRelu;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
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_convscale_reduce<NumDimSpatial,
InDataType,
WeiDataType,
ConvOutDataType,
OutDataType,
ConvElementOp,
ReduceOpId,
InLayout,
WeiLayout,
OutLayout,
3,
AComputeDataType,
BComputeDataType>(
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
......@@ -34,8 +34,17 @@ if (DTYPES)
endif()
message("DTYPES macro set to ${DTYPES}")
else()
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")
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_INT8 "ON")
set(CK_ENABLE_FP16 "ON")
set(CK_ENABLE_FP32 "ON")
set(CK_ENABLE_FP64 "ON")
set(CK_ENABLE_BF16 "ON")
if (GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_ENABLE_FP8 -DCK_ENABLE_BF8)
set(CK_ENABLE_FP8 "ON")
set(CK_ENABLE_BF8 "ON")
endif()
endif()
if (GPU_TARGETS)
......
cmake_minimum_required(VERSION 3.16)
project(composable_kernel_host LANGUAGES CXX HIP)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
......@@ -8,17 +5,9 @@ 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}/..)
find_package(ROCM)
include(ROCMInstallTargets)
include(ROCMTest)
add_compile_options(-std=c++17)
find_package(hip)
## HIP
set(CMAKE_HIP_PLATFORM amd)
set(CMAKE_HIP_COMPILER ${CMAKE_CXX_COMPILER})
set(CMAKE_HIP_EXTENSIONS ON)
message("CMAKE_HIP_COMPILER: ${CMAKE_HIP_COMPILER}")
add_custom_target(codegen)
# add include directories
include_directories(BEFORE
......@@ -32,11 +21,14 @@ list(APPEND CMAKE_MODULE_PATH ${CK_ROOT}/cmake)
include(Embed)
file(GLOB_RECURSE KERNEL_FILES CONFIGURE_DEPENDS
${CK_ROOT}/include/ck/*.hpp)
message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
message(STATUS "RELATIVE: ${CK_ROOT}/include")
#printouts fot debug purposes
#message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
#message(STATUS "RELATIVE: ${CK_ROOT}/include")
add_embed_library(ck_headers ${KERNEL_FILES} RELATIVE ${CK_ROOT}/include)
file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
##message(STATUS "SOURCE_FILES: ${SOURCES}")
# TODO: Use object library
add_library(ck_host STATIC ${SOURCES})
target_link_libraries(ck_host PRIVATE ck_headers)
......@@ -58,6 +50,4 @@ rocm_install(
)
rocm_install(DIRECTORY include/ck DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
if(BUILD_TESTING)
add_subdirectory(test)
endif()
......@@ -76,8 +76,11 @@ std::string SequenceStr(const std::vector<int>& v);
std::string MakeTuple(const std::vector<std::string>& v);
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wglobal-constructors"
template <int... xs>
const std::string S = SequenceStr({xs...});
#pragma clang diagnostic pop
constexpr const char* PassThrough = "ck::tensor_operation::element_wise::PassThrough";
constexpr const char* Bilinear = "ck::tensor_operation::element_wise::Bilinear";
......
......@@ -3,6 +3,7 @@
#include "ck/host/device_gemm_multiple_d/operation.hpp"
#include "ck/host/stringutils.hpp"
#include "ck/host/types.hpp"
#include "ck/host/utils.hpp"
#include <cassert>
......@@ -32,11 +33,11 @@ static std::string GetGemmSpec(const std::size_t m,
}
// function to update prologue/epilogue with user provided operation
void Operation_Xdl_CShuffle::update_prologue(const std::string& prologue)
void Operation_Xdl_CShuffle::update_prologue(const std::string& pro)
{
if(!prologue.empty())
if(!pro.empty())
{
this->prologue = prologue;
this->prologue = pro;
this->cde_elem_op = "CDEElementOp";
}
else
......@@ -45,11 +46,11 @@ void Operation_Xdl_CShuffle::update_prologue(const std::string& prologue)
}
}
void Operation_Xdl_CShuffle::update_epilogue(const std::string& epilogue)
void Operation_Xdl_CShuffle::update_epilogue(const std::string& epi)
{
if(!epilogue.empty())
if(!epi.empty())
{
this->epilogue = epilogue;
this->epilogue = epi;
this->cde_elem_op = "CDEElementOp";
}
else
......
......@@ -4,6 +4,7 @@
#include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
#include <iostream>
#include "ck/host/stringutils.hpp"
#include "ck/host/types.hpp"
#include "ck/host/utils.hpp"
#include <cassert>
......@@ -11,34 +12,15 @@ namespace ck {
namespace host {
namespace conv {
// calculate appropriate Gemm Specification based on input tensor dimensions
// NOTE: in CK, MNKPadding is always used for forward convolution
static std::string GetGemmSpec(const std::size_t m,
const std::size_t n,
const std::size_t k,
const std::size_t m_per_block,
const std::size_t n_per_block,
const std::size_t k_per_block)
{
std::string spec = "";
if(integer_divide_ceil(m, m_per_block) * m_per_block - m != 0)
spec += "M";
if(integer_divide_ceil(n, n_per_block) * n_per_block - n != 0)
spec += "N";
if(integer_divide_ceil(k, k_per_block) * k_per_block - k != 0)
spec += "K";
if(spec == "")
return "ck::tensor_operation::device::GemmSpecialization::Default";
return "ck::tensor_operation::device::GemmSpecialization::" + spec + "Padding";
}
// NOTE: in CK, MNKPadding is always used for forward convolution, so didn't
// add GemmSpec function here
// function to update prologue/epilogue with user provided operation
void Operation_Conv_Fwd_Xdl_Cshuffle::update_prologue(const std::string& prologue)
void Operation_Conv_Fwd_Xdl_Cshuffle::update_prologue(const std::string& pro)
{
if(!prologue.empty())
if(!pro.empty())
{
this->prologue = prologue;
this->prologue = pro;
this->cde_elem_op = "CDEElementOp";
}
else
......@@ -47,11 +29,11 @@ void Operation_Conv_Fwd_Xdl_Cshuffle::update_prologue(const std::string& prologu
}
}
void Operation_Conv_Fwd_Xdl_Cshuffle::update_epilogue(const std::string& epilogue)
void Operation_Conv_Fwd_Xdl_Cshuffle::update_epilogue(const std::string& epi)
{
if(!epilogue.empty())
if(!epi.empty())
{
this->epilogue = epilogue;
this->epilogue = epi;
this->cde_elem_op = "CDEElementOp";
}
else
......@@ -233,6 +215,12 @@ extern "C" __global__ void run_${name}(
${BElementwiseOperation}{},
${CDEElementwiseOperation}{1.0f, 1.0f});
if(!DeviceConv::IsSupportedArgument(arg))
{
printf("Arguement is not supported.\n");
return;
};
constexpr ck::LoopScheduler LoopSched = ck::make_default_loop_scheduler();
// GridwiseGemm
......
......@@ -4,7 +4,10 @@
namespace ck {
namespace host {
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wglobal-constructors"
const std::string config_header = "";
#pragma clang diagnostic pop
std::unordered_map<std::string_view, std::string_view> GetHeaders()
{
......
list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
add_subdirectory(rtc)
file(GLOB TEST_SRCS CONFIGURE_DEPENDS *.cpp)
foreach(TEST_SRC ${TEST_SRCS})
set_source_files_properties(${TEST_SRC} PROPERTIES LANGUAGE HIP)
get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
rocm_add_test_executable(test_host_${BASE_NAME} ${TEST_SRC})
target_link_libraries(test_host_${BASE_NAME} ck_rtc ck_host)
# target_link_libraries(test_host_${BASE_NAME} ${CK_ROOT}/build/lib/libutility.a)
target_include_directories(test_host_${BASE_NAME} PUBLIC include())
target_include_directories(test_host_${BASE_NAME} PUBLIC ${CK_ROOT}/include)
target_include_directories(test_host_${BASE_NAME} PUBLIC ${CK_ROOT}/library/include)
endforeach()
if(NOT INSTANCES_ONLY)
foreach(TEST_SRC ${TEST_SRCS})
set_source_files_properties(${TEST_SRC} PROPERTIES LANGUAGE HIP)
get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
add_executable(codegen_test_${BASE_NAME} ${TEST_SRC})
add_dependencies(codegen codegen_test_${BASE_NAME})
add_dependencies(tests codegen_test_${BASE_NAME})
add_dependencies(check codegen_test_${BASE_NAME})
add_test(NAME codegen_test_${BASE_NAME} COMMAND codegen_test_${BASE_NAME})
message("adding test codegen_test_${BASE_NAME}")
target_link_libraries(codegen_test_${BASE_NAME} ck_rtc ck_host)
target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/codegen/test/include)
target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/include)
target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/library/include)
endforeach()
endif()
......@@ -92,7 +92,6 @@ struct Epilogue
static_cast<int>(prob.C),
static_cast<int>(prob.Y),
static_cast<int>(prob.X)};
ck::Array<ck::index_t, 5> d_lengths = {};
ck::Array<ck::index_t, 5> in_strides{static_cast<int>(prob.C),
static_cast<int>(prob.Hi * prob.Wi * prob.G * prob.C),
......@@ -109,7 +108,6 @@ struct Epilogue
1,
static_cast<int>(prob.X * prob.C),
static_cast<int>(prob.C)};
ck::Array<ck::index_t, 5> d_strides = {};
ck::Array<ck::index_t, 2> conv_filter_strides = {2, 2};
ck::Array<ck::index_t, 2> conv_filter_dilations = {1, 1};
......
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