Skip to content

GitLab

Commit a3b86965 authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into lds_bypass_spilling
parents bdd0f64e fe96e8fb
Hide whitespace changes
Inline Side-by-side
Showing with 813 additions and 98 deletions
CHANGELOG.md
View file @ a3b86965
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
Full documentation for Composable Kernel is not yet available. Full documentation for Composable Kernel is not yet available.
## CK 0.1.1 for ROCm 5.5.0 ## CK 0.2.0 for ROCm 5.5.0
### Fixed ### Fixed
- Fixed a bug in 6-dimensional kernels (#555). - Fixed a bug in 6-dimensional kernels (#555).
...@@ -12,6 +12,7 @@ Full documentation for Composable Kernel is not yet available. ...@@ -12,6 +12,7 @@ Full documentation for Composable Kernel is not yet available.
- Improve proformance of normalization kernel - Improve proformance of normalization kernel
### Added ### Added
- Added support on NAVI3x.
- Added user tutorial (#563). - Added user tutorial (#563).
- Added more instances for irregular GEMM sizes (#560). - Added more instances for irregular GEMM sizes (#560).
- Added inter-wave consumer-producer programming model for GEMM kernels (#310). - Added inter-wave consumer-producer programming model for GEMM kernels (#310).
......
Dockerfile
View file @ a3b86965
...@@ -7,6 +7,8 @@ ARG compiler_commit="" ...@@ -7,6 +7,8 @@ ARG compiler_commit=""
RUN set -xe RUN set -xe
ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/ ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
RUN useradd -rm -d /home/manitera -s /bin/bash -u 1002 manitera
# Add rocm repository # Add rocm repository
RUN apt-get update RUN apt-get update
RUN apt-get install -y wget gnupg RUN apt-get install -y wget gnupg
...@@ -37,6 +39,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -37,6 +39,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
python-dev \ python-dev \
python3-dev \ python3-dev \
python3-pip \ python3-pip \
sshpass \
software-properties-common \ software-properties-common \
rocm-dev \ rocm-dev \
rocm-device-libs \ rocm-device-libs \
......
Jenkinsfile
View file @ a3b86965
...@@ -14,7 +14,6 @@ def show_node_info() { ...@@ -14,7 +14,6 @@ def show_node_info() {
def runShell(String command){ def runShell(String command){
def responseCode = sh returnStatus: true, script: "${command} > tmp.txt" def responseCode = sh returnStatus: true, script: "${command} > tmp.txt"
def output = readFile(file: "tmp.txt") def output = readFile(file: "tmp.txt")
echo "tmp.txt contents: $output"
return (output != "") return (output != "")
} }
...@@ -172,7 +171,7 @@ def cmake_build(Map conf=[:]){ ...@@ -172,7 +171,7 @@ def cmake_build(Map conf=[:]){
if(conf.get("build_install","") == "true") if(conf.get("build_install","") == "true")
{ {
config_targets = 'install ' + config_targets config_targets = 'install ' + config_targets
setup_args = ' -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install' + setup_args setup_args = ' -DBUILD_DEV=On -DCMAKE_INSTALL_PREFIX=../install' + setup_args
} else{ } else{
setup_args = ' -DBUILD_DEV=On' + setup_args setup_args = ' -DBUILD_DEV=On' + setup_args
} }
...@@ -427,6 +426,7 @@ def Build_CK(Map conf=[:]){ ...@@ -427,6 +426,7 @@ def Build_CK(Map conf=[:]){
def variant = env.STAGE_NAME def variant = env.STAGE_NAME
def retimage def retimage
def navi_node = 0
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') { gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
try { try {
...@@ -440,6 +440,9 @@ def Build_CK(Map conf=[:]){ ...@@ -440,6 +440,9 @@ def Build_CK(Map conf=[:]){
else{ else{
echo "GPU is OK" echo "GPU is OK"
} }
if ( runShell('grep -n "gfx1030" clinfo.log') ){
navi_node = 1
}
} }
} }
} }
...@@ -458,6 +461,9 @@ def Build_CK(Map conf=[:]){ ...@@ -458,6 +461,9 @@ def Build_CK(Map conf=[:]){
else{ else{
echo "GPU is OK" echo "GPU is OK"
} }
if ( runShell('grep -n "gfx1030" clinfo.log') ){
navi_node = 1
}
} }
} }
} }
...@@ -466,16 +472,20 @@ def Build_CK(Map conf=[:]){ ...@@ -466,16 +472,20 @@ def Build_CK(Map conf=[:]){
{ {
cmake_build(conf) cmake_build(conf)
dir("build"){ dir("build"){
//run tests and examples if (navi_node == 0 ){
sh 'make -j check' //run tests and examples on all nodes except Navi
//we only need the ckProfiler to run the performance tests, so we pack and stash it sh 'make -j check'
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler' //we only need the ckProfiler to run the performance tests, so we pack and stash it
stash "ckProfiler.tar.gz" sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash "ckProfiler.tar.gz"
}
if (params.RUN_FULL_QA){ if (params.RUN_FULL_QA){
// build deb packages // build deb packages
sh 'make -j package' sh 'make -j package'
archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb' archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb'
archiveArtifacts artifacts: 'composablekernel-tests_*.deb' archiveArtifacts artifacts: 'composablekernel-tests_*.deb'
sh 'mv composablekernel-ckprofiler_*.deb ckprofiler_0.2.0_amd64.deb'
stash "ckprofiler_0.2.0_amd64.deb"
} }
} }
} }
...@@ -543,6 +553,8 @@ def process_results(Map conf=[:]){ ...@@ -543,6 +553,8 @@ def process_results(Map conf=[:]){
unstash "perf_splitK_gemm.log" unstash "perf_splitK_gemm.log"
unstash "perf_onnx_gemm.log" unstash "perf_onnx_gemm.log"
sh "./process_qa_data.sh" sh "./process_qa_data.sh"
unstash "ckprofiler_0.2.0_amd64.deb"
sh "sshpass -p ${env.ck_deb_pw} scp -o StrictHostKeyChecking=no ckprofiler_0.2.0_amd64.deb ${env.ck_deb_user}@${env.ck_deb_ip}:/var/www/html/composable_kernel/"
} }
else{ else{
// unstash perf files to master // unstash perf files to master
...@@ -564,7 +576,7 @@ def process_results(Map conf=[:]){ ...@@ -564,7 +576,7 @@ 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 CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true
0 21 * * * % RUN_FULL_QA=false;COMPILER_VERSION=release;COMPILER_COMMIT= 0 21 * * * % COMPILER_VERSION=release;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : "" 0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
pipeline { pipeline {
...@@ -653,12 +665,28 @@ pipeline { ...@@ -653,12 +665,28 @@ pipeline {
{ {
parallel parallel
{ {
stage("Build CK and run Tests") stage("Build CK and run Tests on MI100/MI200")
{ {
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3 " """ }" setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" """
execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }" execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && 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')
}
}
stage("Build CK and run Tests on Navi")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("navi21") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030;gfx1100;gfx1101;gfx1102" -D CMAKE_CXX_COMPILER="${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')
...@@ -671,7 +699,7 @@ pipeline { ...@@ -671,7 +699,7 @@ pipeline {
{ {
parallel parallel
{ {
stage("Run ckProfiler: gfx908 or gfx90a") stage("Run ckProfiler: gfx90*")
{ {
when { when {
beforeAgent true beforeAgent true
...@@ -680,7 +708,7 @@ pipeline { ...@@ -680,7 +708,7 @@ pipeline {
options { retry(2) } options { retry(2) }
agent{ label rocmnode("gfx908 || gfx90a")} agent{ label rocmnode("gfx908 || gfx90a")}
environment{ environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}" setup_args = """ -DGPU_TARGETS="gfx908;gfx90a" -DBUILD_DEV=On """
} }
steps{ steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release') runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
...@@ -695,7 +723,7 @@ pipeline { ...@@ -695,7 +723,7 @@ pipeline {
options { retry(2) } options { retry(2) }
agent{ label rocmnode("gfx90a")} agent{ label rocmnode("gfx90a")}
environment{ environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}" setup_args = """ -DGPU_TARGETS="gfx90a" -DBUILD_DEV=On """
} }
steps{ steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release') runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......
client_example/09_quantization/CMakeLists.txt
View file @ a3b86965
...@@ -9,3 +9,6 @@ target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composab ...@@ -9,3 +9,6 @@ target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composab
add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp) add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_quantization gemm_quantization.cpp)
target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_operations)
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
View file @ a3b86965
...@@ -28,16 +28,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Cla ...@@ -28,16 +28,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Cla
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
SimpleDeviceMem() = delete; SimpleDeviceMem() = delete;
...@@ -64,8 +63,8 @@ int main(int argc, char* argv[]) ...@@ -64,8 +63,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo}; std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -136,10 +135,11 @@ int main(int argc, char* argv[]) ...@@ -136,10 +135,11 @@ int main(int argc, char* argv[])
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X; std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C + std::size_t num_bytes =
G * sizeof(WeiDataType) * K * Y * X * C + G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K; G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -162,11 +162,12 @@ int main(int argc, char* argv[]) ...@@ -162,11 +162,12 @@ int main(int argc, char* argv[])
} }
} }
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 // run the best intance
if(best_op_id != -1)
{ {
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;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
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;
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ a3b86965
...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clam ...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clam
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -60,8 +60,8 @@ int main(int argc, char* argv[]) ...@@ -60,8 +60,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C}; std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo}; std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -130,10 +130,10 @@ int main(int argc, char* argv[]) ...@@ -130,10 +130,10 @@ int main(int argc, char* argv[])
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X; std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C + std::size_t num_bytes =
G * sizeof(WeiDataType) * K * Y * X * C + G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K; G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -156,11 +156,12 @@ int main(int argc, char* argv[]) ...@@ -156,11 +156,12 @@ int main(int argc, char* argv[])
} }
} }
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 // run the best intance
if(best_op_id != -1)
{ {
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;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
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;
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
View file @ a3b86965
...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_C ...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_C
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -60,8 +60,8 @@ int main(int argc, char* argv[]) ...@@ -60,8 +60,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C}; std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo}; std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -130,10 +130,10 @@ int main(int argc, char* argv[]) ...@@ -130,10 +130,10 @@ int main(int argc, char* argv[])
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X; std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C + std::size_t num_bytes =
G * sizeof(WeiDataType) * K * Y * X * C + G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K; G * sizeof(RequantScaleDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -156,11 +156,12 @@ int main(int argc, char* argv[]) ...@@ -156,11 +156,12 @@ int main(int argc, char* argv[])
} }
} }
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 // run the best intance
if(best_op_id != -1)
{ {
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;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
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;
......
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ a3b86965
...@@ -24,15 +24,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Ac ...@@ -24,15 +24,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Ac
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -56,8 +56,8 @@ int main(int argc, char* argv[]) ...@@ -56,8 +56,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C}; std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X}; std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C}; std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -150,11 +150,11 @@ int main(int argc, char* argv[]) ...@@ -150,11 +150,11 @@ int main(int argc, char* argv[])
} }
} }
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops if(best_op_id != -1)
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{ {
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;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
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;
......
client_example/09_quantization/gemm_quantization.cpp 0 → 100644
View file @ a3b86965
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp"
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using ActivationOp = PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
using ADataType = int8_t;
using BDataType = int8_t;
using EDataType = int8_t;
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
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(int argc, char* argv[])
{
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideE = 1024;
float requant_scale = 0.03;
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<>,
ELayout,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// 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;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
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(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_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});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_bytes =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
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::cout << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id != -1)
{
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;
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(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_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 0;
}
\ No newline at end of file
client_example/17_grouped_gemm_fastgelu/CMakeLists.txt 0 → 100644
View file @ a3b86965
add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_operations)
\ No newline at end of file
client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp 0 → 100644
View file @ a3b86965
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include <random>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using ADataType = F16;
using BDataType = F16;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = FastGelu;
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::mt19937 gen(19391);
std::uniform_int_distribution<> distrib(1, 10);
int group_count = distrib(gen);
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
for(int i = 0; i < group_count; ++i)
{
Ms.push_back(256 + 256 * distrib(gen));
Ns.push_back(256 + 256 * distrib(gen));
Ks.push_back(128 + 128 * distrib(gen));
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]);
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
a_dev_bufs.reserve(group_count);
b_dev_bufs.reserve(group_count);
e_dev_bufs.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_e;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_e.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; ++i)
{
a_dev_bufs.emplace_back(sizeof(ADataType) *
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
b_dev_bufs.emplace_back(sizeof(BDataType) *
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
e_dev_bufs.emplace_back(sizeof(EDataType) *
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideEs[i], {}});
p_a.push_back(a_dev_bufs[i].GetDeviceBuffer());
p_b.push_back(b_dev_bufs[i].GetDeviceBuffer());
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// 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;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
auto p_ds = std::vector<std::array<const void*, 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(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
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 flop = 0, num_btype = 0;
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
{
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
sizeof(EDataType) * Ms[j] * Ns[j];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
if(found)
{
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(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
cmake/EnableCompilerWarnings.cmake
View file @ a3b86965
...@@ -65,7 +65,8 @@ else() ...@@ -65,7 +65,8 @@ else()
-Wuninitialized -Wuninitialized
-Wunreachable-code -Wunreachable-code
-Wunused -Wunused
-Wno-reserved-identifier
-Werror
-Wsign-compare -Wsign-compare
-Wno-extra-semi-stmt -Wno-extra-semi-stmt
) )
......
docs/Doxyfile
View file @ a3b86965
...@@ -775,8 +775,10 @@ WARN_LOGFILE = ...@@ -775,8 +775,10 @@ WARN_LOGFILE =
# spaces. See also FILE_PATTERNS and EXTENSION_MAPPING # spaces. See also FILE_PATTERNS and EXTENSION_MAPPING
# Note: If this tag is empty the current directory is searched. # Note: If this tag is empty the current directory is searched.
INPUT = ../library/include \ INPUT = ../include/ck/tensor_operation/gpu/grid \
../library/include/internal ../include/ck/tensor_operation/gpu/block \
../include/ck/tensor_operation/gpu/thread \
../library/include/ck/library/utility
# This tag can be used to specify the character encoding of the source files # This tag can be used to specify the character encoding of the source files
# that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
...@@ -845,7 +847,7 @@ FILE_PATTERNS = *.c \ ...@@ -845,7 +847,7 @@ FILE_PATTERNS = *.c \
# be searched for input files as well. # be searched for input files as well.
# The default value is: NO. # The default value is: NO.
RECURSIVE = NO RECURSIVE = YES
# The EXCLUDE tag can be used to specify files and/or directories that should be # The EXCLUDE tag can be used to specify files and/or directories that should be
# excluded from the INPUT source files. This way you can easily exclude a # excluded from the INPUT source files. This way you can easily exclude a
......
docs/source/API_Reference_Guide.rst
View file @ a3b86965
=================== *******************
API Reference Guide API Reference Guide
=================== *******************
------------ =================
Introduction Introduction
------------ =================
This document contains details of the APIs for the Composable Kernel (CK) library and introduces some of the key design This document contains details of the APIs for the Composable Kernel (CK) library and introduces some of the key design
principles that are used to write new classes that extend CK functionality. principles that are used to write new classes that extend CK functionality.
...@@ -16,8 +16,37 @@ Using CK API ...@@ -16,8 +16,37 @@ Using CK API
This section describes how to use the CK library API. This section describes how to use the CK library API.
----------------- =================
CK Datatypes CK Datatypes
=================
-----------------
DeviceMem
----------------- -----------------
[TODO] .. doxygenstruct:: DeviceMem
\ No newline at end of file
---------------------------
Kernels For Flashattention
---------------------------
The Flashattention algorithm is defined in :cite:t:`dao2022flashattention`. This sections lists the classes that are
used in the CK GPU implementation of Flashattention.
**Gridwise classes**
.. doxygenstruct:: ck::GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
**Blockwise classes**
.. doxygenstruct:: ck::ThreadGroupTensorSliceTransfer_v4r1
.. doxygenstruct:: ck::BlockwiseGemmXdlops_v2
.. doxygenstruct:: ck::BlockwiseSoftmax
**Threadwise classes**
.. doxygenstruct:: ck::ThreadwiseTensorSliceTransfer_StaticToStatic
.. bibliography::
\ No newline at end of file
docs/source/conf.py
View file @ a3b86965
...@@ -59,10 +59,13 @@ if read_the_docs_build: ...@@ -59,10 +59,13 @@ if read_the_docs_build:
# Add any Sphinx extension module names here, as strings. They can be # Add any Sphinx extension module names here, as strings. They can be
# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
# ones. # ones.
extensions = ['sphinx.ext.mathjax', 'breathe'] extensions = ['sphinx.ext.mathjax', 'breathe', 'sphinxcontrib.bibtex']
breathe_projects = { "CK": "../docBin/xml" } breathe_projects = { "CK": "../docBin/xml" }
breathe_default_project = "CK" breathe_default_project = "CK"
bibtex_bibfiles = ['refs.bib']
# Add any paths that contain templates here, relative to this directory. # Add any paths that contain templates here, relative to this directory.
templates_path = ['_templates'] templates_path = ['_templates']
......
docs/source/refs.bib 0 → 100644
View file @ a3b86965
@article{dao2022flashattention,
title={Flashattention: Fast and memory-efficient exact attention with io-awareness},
author={Dao, Tri and Fu, Daniel Y and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
journal={arXiv preprint arXiv:2205.14135},
year={2022}
}
example/01_gemm/CMakeLists.txt
View file @ a3b86965
...@@ -38,7 +38,7 @@ add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp) ...@@ -38,7 +38,7 @@ add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16) add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp64) add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_custom_target(example_gemm_wmma) add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp) add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_dependencies(example_gemm_wmma example_gemm_wmma_fp16) add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
......
example/02_gemm_bilinear/CMakeLists.txt
View file @ a3b86965
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp) add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp) add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
endif() endif()
example/14_gemm_quantization/CMakeLists.txt
View file @ a3b86965
# dlops
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
# xdlops
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp) add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp) add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
\ No newline at end of file
example/14_gemm_quantization/gemm_dl_quantization_int8.cpp 0 → 100644
View file @ a3b86965
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using I8 = int8_t;
using I32 = int32_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using ActivationOp = PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
using ADataType = I8;
using BDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I32;
using DsDataType = ck::Tuple<>;
using EDataType = I8;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Dl<
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
AccDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmDefault,
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
16, // K0PerBlock
4, // K1
4, // M1PerThread
4, // N1PerThread
1, // KPerThread
S<8, 2>, // M1N1ThreadClusterM1Xs
S<8, 2>, // M1N1ThreadClusterN1Xs
S<8, 1, 1, 4>, // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
S<2, 1, 128, 1>, // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
S<1, 2, 0, 3>, // ABlockTransferThreadClusterArrangeOrder
S<1, 2, 0, 3>, // ABlockTransferSrcAccessOrder
S<4, 1, 1, 4>, // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
S<1, 2, 0, 3>, // ABlockTransferSrcVectorTensorContiguousDimOrder
S<1, 1, 1, 4>, // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
S<8, 1, 1, 4>, // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
S<2, 1, 128, 1>, // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
S<1, 2, 0, 3>, // BBlockTransferThreadClusterArrangeOrder
S<1, 2, 0, 3>, // BBlockTransferSrcAccessOrder
S<4, 1, 1, 4>, // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
S<1, 2, 0, 3>, // BBlockTransferSrcVectorTensorContiguousDimOrder
S<1, 1, 1, 4>, // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
5, // CThreadTransferSrcDstVectorDim
4>; // CThreadTransferDstScalarPerVector
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
int main()
{
bool do_verification = true;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideE = 1024;
float requant_scale = 0.03;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1_uz}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1_uz, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument);
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
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