Skip to content

GitLab

Commit 8a370fbb authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into group_conv

parents d8fdd226 85978e02
Show whitespace changes
Inline Side-by-side
Showing with 2721 additions and 213 deletions
Jenkinsfile
View file @ 8a370fbb
......@@ -11,6 +11,12 @@ def show_node_info() {
"""
}
def runShell(String command){
def responseCode = sh returnStatus: true, script: "${command} &> tmp.txt"
def output = readFile(file: "tmp.txt")
return (output != "")
}
def cmake_build(Map conf=[:]){
def compiler = conf.get("compiler","/opt/rocm/bin/hipcc")
......@@ -60,7 +66,7 @@ def cmake_build(Map conf=[:]){
"""
def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
// reduce parallelism when compiling, clang uses too much memory
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(( \$(nproc) / 1 )) ${config_targets}")
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(( \$(nproc) / 2 )) ${config_targets}")
def execute_cmd = conf.get("execute_cmd", "")
def cmd = conf.get("cmd", """
......@@ -113,7 +119,14 @@ def buildHipClangJob(Map conf=[:]){
retimage = docker.build("${image}", dockerArgs + '.')
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
echo "GPU not found"
throw e
}
else{
echo "GPU is OK"
}
}
}
}
......@@ -125,7 +138,14 @@ def buildHipClangJob(Map conf=[:]){
retimage = docker.build("${image}", dockerArgs + " --no-cache .")
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo |tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
echo "GPU not found"
throw e
}
else{
echo "GPU is OK"
}
}
}
}
......@@ -133,7 +153,14 @@ def buildHipClangJob(Map conf=[:]){
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 5, unit: 'HOURS')
{
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
echo "GPU not found"
throw e
}
else{
echo "GPU is OK"
}
cmake_build(conf)
}
}
......@@ -145,7 +172,6 @@ def reboot(){
build job: 'reboot-slaves', propagate: false , parameters: [string(name: 'server', value: "${env.NODE_NAME}"),]
}
def buildHipClangJobAndReboot(Map conf=[:]){
try{
buildHipClangJob(conf)
......@@ -162,7 +188,6 @@ def buildHipClangJobAndReboot(Map conf=[:]){
}
}
def runCKProfiler(Map conf=[:]){
show_node_info()
......@@ -189,7 +214,6 @@ def runCKProfiler(Map conf=[:]){
}
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
......@@ -197,7 +221,14 @@ def runCKProfiler(Map conf=[:]){
retimage = docker.build("${image}", dockerArgs + '.')
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
echo "GPU not found"
throw e
}
else{
echo "GPU is OK"
}
}
}
}
......@@ -209,89 +240,69 @@ def runCKProfiler(Map conf=[:]){
retimage = docker.build("${image}", dockerArgs + " --no-cache .")
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo'
sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
echo "GPU not found"
throw e
}
else{
echo "GPU is OK"
}
}
}
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 5, unit: 'HOURS')
timeout(time: 24, unit: 'HOURS')
{
cmake_build(conf)
dir("script"){
//run gemm performance tests
def gemm_log = "perf_gemm_${gpu_arch}.log"
sh "rm -f ${gemm_log}"
sh "echo Branch name: ${env.BRANCH_NAME} > ${gemm_log}"
sh "echo Node name: ${NODE_NAME} >> ${gemm_log}"
sh "echo GPU_arch name: ${gpu_arch} >> ${gemm_log}"
sh "rocminfo | grep 'Compute Unit:' >> ${gemm_log} "
sh "hipcc --version | grep -e 'HIP version' >> ${gemm_log}"
if (params.RUN_FULL_QA){
def qa_log = "qa_${gpu_arch}.log"
if (params.USE_9110){
sh "echo Environment type: CI_9110 >> ${gemm_log}"
sh "./run_full_performance_tests.sh 1 QA_9110 ${gpu_arch} ${env.BRANCH_NAME} ${NODE_NAME}"
}
else{
sh "echo Environment type: CI_release >> ${gemm_log}"
}
sh "/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}"
sh "./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 0 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 1 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 0 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 2 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 0 3 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a ${gemm_log}"
sh "./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a ${gemm_log}"
//results will be parsed, stored, and analyzed within the python script
//the script will return 0 if the performance criteria are met
//or return 1 if the criteria are not met
archiveArtifacts "${gemm_log}"
sh "python3 process_perf_data.py ${gemm_log} "
//run resnet50 test
def resnet256_log = "perf_resnet50_N256_${gpu_arch}.log"
sh "rm -f ${resnet256_log}"
sh "echo Branch name: ${env.BRANCH_NAME} > ${resnet256_log}"
sh "echo Node name: ${NODE_NAME} >> ${resnet256_log}"
sh "echo GPU_arch name: ${gpu_arch} >> ${resnet256_log}"
sh "rocminfo | grep 'Compute Unit:' >> ${resnet256_log} "
sh "hipcc --version | grep -e 'HIP version' >> ${resnet256_log}"
if (params.USE_9110){
sh "echo Environment type: CI_9110 >> ${resnet256_log}"
sh "./run_full_performance_tests.sh 1 QA_release ${gpu_arch} ${env.BRANCH_NAME} ${NODE_NAME}"
}
archiveArtifacts "perf_gemm_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N256_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N4_${gpu_arch}.log"
archiveArtifacts "perf_bathced_gemm_${gpu_arch}.log"
archiveArtifacts "perf_grouped_gemm_${gpu_arch}.log"
archiveArtifacts "perf_fwd_conv_${gpu_arch}.log"
archiveArtifacts "perf_bwd_conv_${gpu_arch}.log"
archiveArtifacts "perf_fusion_${gpu_arch}.log"
archiveArtifacts "perf_reduction_${gpu_arch}.log"
// stash perf files to master
stash name: "perf_gemm_${gpu_arch}.log"
stash name: "perf_resnet50_N256_${gpu_arch}.log"
stash name: "perf_resnet50_N4_${gpu_arch}.log"
stash name: "perf_bathced_gemm_${gpu_arch}.log"
stash name: "perf_grouped_gemm_${gpu_arch}.log"
stash name: "perf_fwd_conv_${gpu_arch}.log"
stash name: "perf_bwd_conv_${gpu_arch}.log"
stash name: "perf_fusion_${gpu_arch}.log"
stash name: "perf_reduction_${gpu_arch}.log"
//we will process results on the master node
}
else{
sh "echo Environment type: CI_release >> ${resnet256_log}"
}
sh "/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet256_log}"
//first run tests with N=256
sh "./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a ${resnet256_log}"
archiveArtifacts "${resnet256_log}"
sh "python3 process_perf_data.py ${resnet256_log} "
//then run with N=4
def resnet4_log = "perf_resnet50_N4_${gpu_arch}.log"
sh "rm -f ${resnet4_log}"
sh "echo Branch name: ${env.BRANCH_NAME} > ${resnet4_log}"
sh "echo Node name: ${NODE_NAME} >> ${resnet4_log}"
sh "echo GPU_arch name: ${gpu_arch} >> ${resnet4_log}"
sh "rocminfo | grep 'Compute Unit:' >> ${resnet4_log} "
sh "hipcc --version | grep -e 'HIP version' >> ${resnet4_log}"
if (params.USE_9110){
sh "echo Environment type: CI_9110 >> ${resnet4_log}"
sh "./run_performance_tests.sh 0 CI_9110 ${gpu_arch} ${env.BRANCH_NAME} ${NODE_NAME}"
}
else{
sh "echo Environment type: CI_release >> ${resnet4_log}"
sh "./run_performance_tests.sh 0 CI_release ${gpu_arch} ${env.BRANCH_NAME} ${NODE_NAME}"
}
sh "/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet4_log}"
sh "./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a ${resnet4_log}"
archiveArtifacts "${resnet4_log}"
sh "python3 process_perf_data.py ${resnet4_log} "
archiveArtifacts "perf_gemm_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N256_${gpu_arch}.log"
archiveArtifacts "perf_resnet50_N4_${gpu_arch}.log"
// stash perf files to master
stash name: "perf_gemm_${gpu_arch}.log"
stash name: "perf_resnet50_N256_${gpu_arch}.log"
stash name: "perf_resnet50_N4_${gpu_arch}.log"
//we will process the results on the master node
}
}
}
}
......@@ -299,7 +310,6 @@ def runCKProfiler(Map conf=[:]){
return retimage
}
def runPerfTest(Map conf=[:]){
try{
runCKProfiler(conf)
......@@ -316,8 +326,76 @@ def runPerfTest(Map conf=[:]){
}
}
def process_results(Map conf=[:]){
env.HSA_ENABLE_SDMA=0
checkout scm
def image = "composable_kernels"
def prefixpath = "/opt/rocm"
def gpu_arch = conf.get("gpu_arch", "gfx908")
// Jenkins is complaining about the render group
def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1"
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg GPU_ARCH='${gpu_arch}' --build-arg compiler_version='release' "
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
try {
retimage = docker.build("${image}", dockerArgs + '.')
}
catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
echo "The job was cancelled or aborted"
throw e
}
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 1, unit: 'HOURS'){
try{
dir("script"){
if (params.RUN_FULL_QA){
// unstash perf files to master
unstash "perf_gemm_${gpu_arch}.log"
unstash "perf_resnet50_N256_${gpu_arch}.log"
unstash "perf_resnet50_N4_${gpu_arch}.log"
unstash "perf_bathced_gemm_${gpu_arch}.log"
unstash "perf_grouped_gemm_${gpu_arch}.log"
unstash "perf_fwd_conv_${gpu_arch}.log"
unstash "perf_bwd_conv_${gpu_arch}.log"
unstash "perf_fusion_${gpu_arch}.log"
unstash "perf_reduction_${gpu_arch}.log"
sh "./process_qa_data.sh ${gpu_arch}"
}
else{
// unstash perf files to master
unstash "perf_gemm_${gpu_arch}.log"
unstash "perf_resnet50_N256_${gpu_arch}.log"
unstash "perf_resnet50_N4_${gpu_arch}.log"
sh "./process_perf_data.sh ${gpu_arch}"
}
}
}
catch(e){
echo "throwing error exception while processing performance test results"
echo 'Exception occurred: ' + e.toString()
throw e
}
}
}
}
//launch develop branch daily at 23:00 in FULL_QA mode
//CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;USE_9110=true''' : ""
pipeline {
agent none
//triggers {
// cron(CRON_SETTINGS)
//}
options {
parallelsAlwaysFailFast()
}
......@@ -325,7 +403,11 @@ pipeline {
booleanParam(
name: "USE_9110",
defaultValue: true,
description: "")
description: "Select compiler version: 9110 (default) or release")
booleanParam(
name: "RUN_FULL_QA",
defaultValue: false,
description: "Select whether to run small set of performance tests (default) or full QA")
}
environment{
dbuser = "${dbuser}"
......@@ -438,6 +520,25 @@ pipeline {
}
}
}
stage("Process Performance Test Results")
{
parallel
{
stage("Process results for gfx908"){
agent { label 'mici' }
steps{
process_results(gpu_arch: "gfx908")
}
}
stage("Process results for gfx90a"){
agent { label 'mici' }
steps{
process_results(gpu_arch: "gfx90a")
}
}
}
}
/* enable after the cmake file supports packaging
stage("Packages") {
when {
......
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
View file @ 8a370fbb
......@@ -29,34 +29,40 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = 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 = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// static constexpr auto GemmMNPadding =
// ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| Num|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| |
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1>;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
// clang-format off
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
int main(int argc, char* argv[])
{
......@@ -81,11 +87,11 @@ int main(int argc, char* argv[])
int group_count = rand() % 16 + 1;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
gemm_shapes.reserve(group_count);
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
......@@ -93,7 +99,11 @@ int main(int argc, char* argv[])
int N = 128 + 128 * i;
int K = 64 + 64 * i;
gemm_shapes.push_back({M, N, K, K, K, N});
int stride_A = K;
int stride_B = K;
int stride_C = N;
gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
}
auto f_host_tensor_descriptor =
......@@ -111,10 +121,9 @@ int main(int argc, char* argv[])
};
std::vector<Tensor<ADataType>> a_tensors;
;
std::vector<Tensor<BDataType>> b_tensors;
std::vector<Tensor<CDataType>> c_host_tensors;
std::vector<Tensor<CDataType>> c_device_tensors;
std::vector<Tensor<EDataType>> c_host_tensors;
std::vector<Tensor<EDataType>> c_device_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
......@@ -131,25 +140,25 @@ int main(int argc, char* argv[])
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
c_host_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
c_device_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
<< " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
<< std::endl;
flop += std::size_t(2) * gemm_shapes[i].M * gemm_shapes[i].K * gemm_shapes[i].N;
flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSize();
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
switch(init_method)
{
......@@ -168,14 +177,14 @@ int main(int argc, char* argv[])
}
}
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
......@@ -187,14 +196,16 @@ int main(int argc, char* argv[])
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto c_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
std::vector<std::array<const void*, 0>> p_Ds = {};
// do GEMM
auto argument =
gemm.MakeArgument(p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
auto argument = gemm.MakeArgument(
p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
......@@ -219,7 +230,7 @@ int main(int argc, char* argv[])
bool pass = true;
if(do_verification)
{
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
auto ref_gemm = ReferenceGemmInstance{};
......
example/24_batched_gemm_e_permute/batched_gemm_e_permute_xdl_fp16.cpp
View file @ 8a370fbb
......@@ -26,31 +26,29 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using EDataType = ck::half_t;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using EDataType = F16;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmEPermuteXdl
//######| ALayout| BLayout| AData| BData| AccData| CShuffle| EData| A| B| C| GEMM| Num| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | Type| Type| Type| Data| Type| Elementwise| Elementwise| Elementwise|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | Type| | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// < Row, Col, F16, F16, F32, F16, F16, PassThrough, PassThrough, PassThrough, MNPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8>;
< Row, Col, F16, F16, F32, F16, F16, PassThrough, PassThrough, PassThrough, MNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format off
//######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | Type| Type| Type| DataType| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
......@@ -62,15 +60,18 @@ int main(int argc, char* argv[])
int init_method = 1;
bool time_kernel = false;
const int M = 88;
const int N = 64;
const int K = 88;
const int M = 256;
const int N = 128;
const int K = 64;
const int stride_A = K;
const int stride_B = K;
const int G0 = 1024;
const int G1 = 10;
const int batch_stride_A = M * K;
const int batch_stride_B = K * N;
const int G0 = 16;
const int G1 = 8;
const int batch_count = G0 * G1;
......@@ -102,21 +103,24 @@ int main(int argc, char* argv[])
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({row * stride, stride, 1}));
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({col * stride, 1, stride}));
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(batch_count, M, K, stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(batch_count, K, N, stride_B, BLayout{}));
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
auto f_host_e_tensor_descriptor = [](std::size_t G0_,
std::size_t G1_,
......@@ -169,7 +173,7 @@ int main(int argc, char* argv[])
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
// do GEM
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
......@@ -178,11 +182,13 @@ int main(int argc, char* argv[])
K,
stride_A,
stride_B,
batch_stride_A,
batch_stride_B,
batched_gemm_e_permute_desc,
batch_count,
a_element_op,
b_element_op,
cde_element_op,
batch_count);
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
......
example/28_grouped_gemm_bias/CMakeLists.txt 0 → 100644
View file @ 8a370fbb
add_example_executable(example_grouped_gemm_bias_xdl_fp16 grouped_gemm_bias_xdl_fp16.cpp)
example/28_grouped_gemm_bias/grouped_gemm_bias_xdl_fp16.cpp 0 → 100644
View file @ 8a370fbb
// 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/device_grouped_gemm_xdl.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/utility/check_err.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/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using DsLayout = ck::Tuple<DLayout>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
// clang-format off
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
int group_count = rand() % 16 + 1;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<const void*> p_a, p_b;
std::vector<std::array<const void*, 1>> p_ds;
std::vector<void*> p_c;
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
int M = 256 + 256 * i;
int N = 128 + 128 * i;
int K = 64 + 64 * i;
int stride_A = K;
int stride_B = K;
int stride_C = N;
std::vector<ck::index_t> stride_Ds = {0};
gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, stride_Ds});
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
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}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<BDataType>> b_tensors;
std::vector<Tensor<DDataType>> d_tensors;
std::vector<Tensor<EDataType>> e_host_tensors;
std::vector<Tensor<EDataType>> e_device_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
d_tensors.reserve(group_count);
e_host_tensors.reserve(group_count);
e_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, d_tensors_device,
e_tensors_device;
a_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
d_tensors_device.reserve(group_count);
e_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
d_tensors.push_back(Tensor<DDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_Ds_[0], ELayout{})));
e_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
e_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
<< " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << e_device_tensors[i].mDesc
<< std::endl;
flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
switch(init_method)
{
case 0: break;
case 1:
a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
d_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
}
}
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
d_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(DDataType) * d_tensors[i].mDesc.GetElementSpaceSize()));
e_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
d_tensors_device[i]->ToDevice(d_tensors[i].mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
p_ds.push_back({d_tensors_device[i]->GetDeviceBuffer()});
p_c.push_back(e_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
auto argument = gemm.MakeArgument(
p_a, p_b, p_ds, p_c, gemm_descs, a_element_op, b_element_op, cde_element_op);
DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
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});
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;
bool pass = true;
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
e_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
b_tensors[i],
e_host_tensors[i],
a_element_op,
b_element_op,
PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < gemm_descs[i].M_; ++m)
{
for(int n = 0; n < gemm_descs[i].N_; ++n)
{
cde_element_op(
e_host_tensors[i](m, n), e_host_tensors[i](m, n), d_tensors[i](m, n));
}
}
pass &= ck::utils::check_err(e_device_tensors[i].mData, e_host_tensors[i].mData);
}
}
return pass ? 0 : 1;
}
example/29_batched_gemm_multi_d/CMakeLists.txt 0 → 100644
View file @ 8a370fbb
add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_bias_xdl_fp16 batched_gemm_bias_xdl_fp16.cpp)
example/29_batched_gemm_multi_d/batched_gemm_bias_xdl_fp16.cpp 0 → 100644
View file @ 8a370fbb
#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/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/utility/check_err.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/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using DsLayout = ck::Tuple<DLayout>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
const int M = 256 * (rand() % 16 + 1);
const int N = 128 * (rand() % 16 + 1);
const int K = 64 * (rand() % 16 + 1);
const int stride_A = K;
const int stride_B = K;
const int stride_D = 0;
const int stride_E = N;
const int batch_stride_A = M * K;
const int batch_stride_B = K * N;
const int batch_stride_D = N;
const int batch_stride_E = M * N;
const int batch_count = 16;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
// GEMM shape
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
Tensor<DDataType> d_g_m_n(
f_host_tensor_descriptor(batch_count, M, N, stride_D, batch_stride_D, DLayout{}));
Tensor<EDataType> e_g_m_n_device_result(
f_host_tensor_descriptor(batch_count, M, N, stride_E, batch_stride_E, ELayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "d_g_m_n: " << d_g_m_n.mDesc << std::endl;
std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_g_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_g_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_g_m_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
d_device_buf.ToDevice(d_g_m_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{d_device_buf.GetDeviceBuffer()},
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
batch_count,
stride_A,
stride_B,
{stride_D},
stride_E,
batch_stride_A,
batch_stride_B,
{batch_stride_D},
batch_stride_E,
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) * batch_count * M * N * K;
std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
sizeof(BDataType) * batch_count * K * N +
sizeof(EDataType) * batch_count * 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;
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
EDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
Tensor<EDataType> e_g_m_n_host_result(
f_host_tensor_descriptor(batch_count, M, N, stride_E, batch_stride_E, ELayout{}));
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int g = 0; g < batch_count; g++)
{
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_g_m_n_host_result(g, m, n),
e_g_m_n_host_result(g, m, n),
d_g_m_n(g, m, n));
}
}
}
pass = ck::utils::check_err(
e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
}
return pass ? 0 : 1;
}
example/29_batched_gemm_multi_d/batched_gemm_xdl_fp16.cpp 0 → 100644
View file @ 8a370fbb
#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/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
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 ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = 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 = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
ReferenceBatchedGemm<ADataType, BDataType, EDataType, AElementOp, BElementOp, CDEElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
const int M = 256 * (rand() % 16 + 1);
const int N = 128 * (rand() % 16 + 1);
const int K = 64 * (rand() % 16 + 1);
const int stride_A = K;
const int stride_B = K;
const int stride_C = N;
const int batch_stride_A = M * K;
const int batch_stride_B = K * N;
const int batch_stride_C = M * N;
const int batch_count = 16;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
// GEMM shape
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
Tensor<EDataType> e_g_m_n_device_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
batch_count,
stride_A,
stride_B,
{},
stride_C,
batch_stride_A,
batch_stride_B,
{},
batch_stride_C,
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) * batch_count * M * N * K;
std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
sizeof(BDataType) * batch_count * K * N +
sizeof(EDataType) * batch_count * 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;
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
Tensor<EDataType> e_g_m_n_host_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument);
pass = ck::utils::check_err(
e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
}
return pass ? 0 : 1;
}
example/CMakeLists.txt
View file @ 8a370fbb
......@@ -42,4 +42,6 @@ add_subdirectory(24_batched_gemm_e_permute)
add_subdirectory(25_gemm_bias_e_permute)
add_subdirectory(26_contraction)
add_subdirectory(27_layernorm)
add_subdirectory(28_group_convnd_fwd_bias_relu)
add_subdirectory(28_grouped_gemm_bias)
add_subdirectory(29_batched_gemm_multi_d)
add_subdirectory(30_grouped_convnd_fwd_bias_relu)
include/ck/tensor_operation/gpu/device/device_batched_gemm_c_permute_xdl.hpp 0 → 100644
View file @ 8a370fbb
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_c_permute.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
/*
* \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
*
* \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
* given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
* strided batched, but we can easily extend to other layouts. The returned offset can be either \p
* index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
* limitations.
*
* \tparam Block2CTileMap Block2CTileMap::CalculateBottomIndex() takes in id of a workgroup and
* returns the 2D index of the tile that it computes. \see
* GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
*
* \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
* tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
* descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
* device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
* DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
* pointer offset into \p ComputePtrOffsetOfStridedBatch.
*
* \note \p Block2CTileMap allows customized mapping between a workgroup and the C-tile it computes.
* Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
* realize BatchedGemmCPermute and GroupedGemm (and the corresponding GEMM fusion).
*
*/
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputePtrOffsetOfBatch,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_batched_gemm_c_permute_xdl(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_e_grid,
const index_t batch_count,
const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(
p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
ck::Tuple<>{},
p_e_grid + c_batch_offset,
p_shared,
a_element_op,
b_element_op,
cde_element_op,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ck::StaticallyIndexedArray<
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
0>{},
c_grid_desc_mblock_mperblock_nblock_nperblock,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_e_grid;
ignore = batch_count;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = compute_ptr_offset_of_batch;
ignore = block_2_ctile_map;
#endif
}
template <typename ALayout,
typename BLayout,
typename DELayout,
typename ADataType,
typename BDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmCPermuteXdl : public DeviceBatchedGemmCPermute<ALayout,
BLayout,
DELayout,
ADataType,
BDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
using DeviceOp = DeviceBatchedGemmCPermuteXdl;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto MPad = M - MRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
}
static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
}();
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
assert(K % BK1 == 0);
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
assert(KRaw % BK1 == 0);
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
assert(K % BK1 == 0);
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
assert(KRaw % BK1 == 0);
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
}
static auto
MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t stride_M, index_t stride_N)
{
const auto c_grid_desc_mraw_nraw = [&]() {
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(stride_M, stride_N));
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto MPad = M - MRaw;
const auto NPad = N - NRaw;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
}
static auto MakeEGridDescriptor_G0_G1_M_N(index_t G0,
index_t G1,
index_t MRaw,
index_t NRaw,
index_t stride_G0,
index_t stride_G1,
index_t stride_M,
index_t stride_N)
{
const auto e_grid_desc_g0_g1_mraw_nraw = [&]() {
return make_naive_tensor_descriptor(
make_tuple(G0, G1, MRaw, NRaw),
make_tuple(stride_G0, stride_G1, stride_M, stride_N));
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto MPad = M - MRaw;
const auto NPad = N - NRaw;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(
e_grid_desc_g0_g1_mraw_nraw,
make_tuple(make_pass_through_transform(G0),
make_pass_through_transform(G1),
make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
e_grid_desc_g0_g1_mraw_nraw,
make_tuple(make_pass_through_transform(G0),
make_pass_through_transform(G1),
make_right_pad_transform(MRaw, MPad),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
e_grid_desc_g0_g1_mraw_nraw,
make_tuple(make_pass_through_transform(G0),
make_pass_through_transform(G1),
make_pass_through_transform(MRaw),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
}
else
{
// not pad M or N
return e_grid_desc_g0_g1_mraw_nraw;
}
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1, 1));
using EGridDesc_G0_G1_M_N = decltype(MakeEGridDescriptor_G0_G1_M_N(1, 1, 1, 1, 1, 1, 1, 1));
struct ComputePtrOffsetOfStridedBatch
{
ComputePtrOffsetOfStridedBatch(index_t Batchstride_A,
index_t Batchstride_B,
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n)
: Batchstride_A_(Batchstride_A),
Batchstride_B_(Batchstride_B),
e_grid_desc_g0_g1_m_n_(e_grid_desc_g0_g1_m_n)
{
}
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(Batchstride_A_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(Batchstride_B_);
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
const index_t G1 = e_grid_desc_g0_g1_m_n_.GetLength(I1);
index_t b0 = g_idx / G1;
index_t b1 = g_idx - b0 * G1; // g_idx % G1
return e_grid_desc_g0_g1_m_n_.CalculateOffset(make_multi_index(b0, b1, 0, 0));
}
private:
index_t Batchstride_A_;
index_t Batchstride_B_;
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
};
using GridwiseGemm = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
EGridDesc_M_N,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEBlockTransferScalarPerVector_NPerBlock,
LoopSched>;
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = decltype(
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}));
using Block2CTileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
EDataType* p_e_grid,
index_t M,
index_t N,
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_e_grid_{p_e_grid},
BatchCount_(BatchCount),
a_grid_desc_ak0_m_ak1_{
DeviceBatchedGemmCPermuteXdl::MakeAGridDescriptor_AK0_M_AK1(M, K, stride_A)},
b_grid_desc_bk0_n_bk1_{
DeviceBatchedGemmCPermuteXdl::MakeBGridDescriptor_BK0_N_BK1(K, N, stride_B)},
e_grid_desc_m_n_{DeviceBatchedGemmCPermuteXdl::MakeEGridDescriptor_M_N(
batched_gemm_c_permute_desc.M_,
batched_gemm_c_permute_desc.N_,
batched_gemm_c_permute_desc.stride_M_,
batched_gemm_c_permute_desc.stride_N_)},
e_grid_desc_g0_g1_m_n_{DeviceBatchedGemmCPermuteXdl::MakeEGridDescriptor_G0_G1_M_N(
batched_gemm_c_permute_desc.G0_,
batched_gemm_c_permute_desc.G1_,
batched_gemm_c_permute_desc.M_,
batched_gemm_c_permute_desc.N_,
batched_gemm_c_permute_desc.stride_G0_,
batched_gemm_c_permute_desc.stride_G1_,
batched_gemm_c_permute_desc.stride_M_,
batched_gemm_c_permute_desc.stride_N_)},
c_grid_desc_mblock_mperblock_nblock_nperblock{},
compute_ptr_offset_of_batch_{batch_stride_A, batch_stride_B, e_grid_desc_g0_g1_m_n_},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
e_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
EDataType* p_e_grid_;
index_t BatchCount_;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
EGridDesc_M_N e_grid_desc_m_n_;
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock;
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
Block2CTileMap block_2_ctile_map_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceBatchedGemmCPermuteXdl::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
{
std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.e_grid_desc_m_n_{" << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseBatchedGemmCPermute_km_kn_m0m1n0n1_xdlops_v2r3 has invalid "
"setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.BatchCount_;
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
float ave_time = 0;
auto launch_kernel = [&](auto has_main_k_block_loop_) {
const auto kernel = kernel_batched_gemm_c_permute_xdl<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
EDataType,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
ComputePtrOffsetOfStridedBatch,
remove_reference_t<Block2CTileMap>,
has_main_k_block_loop_>;
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_e_grid_,
arg.BatchCount_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.compute_ptr_offset_of_batch_,
arg.block_2_ctile_map_);
};
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
ave_time = launch_kernel(integral_constant<bool, true>{});
}
else
{
ave_time = launch_kernel(integral_constant<bool, false>{});
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
EDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{p_a,
p_b,
p_c,
M,
N,
K,
stride_A,
stride_B,
batch_stride_A,
batch_stride_B,
batched_gemm_c_permute_desc,
BatchCount,
a_element_op,
b_element_op,
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<EDataType*>(p_c),
M,
N,
K,
stride_A,
stride_B,
batch_stride_A,
batch_stride_B,
batched_gemm_c_permute_desc,
BatchCount,
a_element_op,
b_element_op,
cde_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceBatchedGemmCPermuteXdl"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
View file @ 8a370fbb
......@@ -14,7 +14,13 @@ struct BatchedGemmEPermuteDesc
ck::index_t stride_G0_, stride_G1_, stride_M_, stride_N_;
};
template <typename AElementwiseOperation,
template <typename ALayout,
typename BLayout,
typename DELayout,
typename ADataType,
typename BDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceBatchedGemmEPermute : public BaseOperator
......@@ -28,11 +34,13 @@ struct DeviceBatchedGemmEPermute : public BaseOperator
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
ck::index_t BatchCount) = 0;
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute_xdl.hpp
View file @ 8a370fbb
......@@ -118,6 +118,7 @@ __global__ void
template <typename ALayout,
typename BLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename AccDataType,
......@@ -157,7 +158,13 @@ template <typename ALayout,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementwiseOperation,
struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<ALayout,
BLayout,
ELayout,
ADataType,
BDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
......@@ -389,11 +396,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementw
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
index_t BatchCount)
CDEElementwiseOperation cde_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_e_grid_{p_e_grid},
......@@ -419,10 +428,7 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementw
batched_gemm_e_permute_desc.stride_G1_,
batched_gemm_e_permute_desc.stride_M_,
batched_gemm_e_permute_desc.stride_N_)},
compute_ptr_offset_of_batch_{
type_convert<index_t>(a_grid_desc_ak0_m_ak1_.GetElementSpaceSize()),
type_convert<index_t>(b_grid_desc_bk0_n_bk1_.GetElementSpaceSize()),
e_grid_desc_g0_g1_m_n_},
compute_ptr_offset_of_batch_{batch_stride_A, batch_stride_B, e_grid_desc_g0_g1_m_n_},
block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
......@@ -584,11 +590,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementw
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
index_t BatchCount)
CDEElementwiseOperation cde_element_op)
{
return Argument{p_a,
p_b,
......@@ -598,11 +606,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementw
K,
stride_A,
stride_B,
batch_stride_A,
batch_stride_B,
batched_gemm_e_permute_desc,
BatchCount,
a_element_op,
b_element_op,
cde_element_op,
BatchCount};
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -617,11 +627,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementw
index_t K,
index_t stride_A,
index_t stride_B,
index_t batch_stride_A,
index_t batch_stride_B,
BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
index_t BatchCount,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
index_t BatchCount) override
CDEElementwiseOperation cde_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -631,11 +643,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<AElementw
K,
stride_A,
stride_B,
batch_stride_A,
batch_stride_B,
batched_gemm_e_permute_desc,
BatchCount,
a_element_op,
b_element_op,
cde_element_op,
BatchCount);
cde_element_op);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp 0 → 100644
View file @ 8a370fbb
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceBatchedGemmMultiD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t Batch,
index_t StrideA,
index_t StrideB,
const std::array<ck::index_t, NumDTensor>& StrideDs,
index_t StrideE,
index_t BatchStrideA,
index_t BatchStrideB,
const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
index_t BatchStrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp 0 → 100644
View file @ 8a370fbb
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
/*
* \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
*
* \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
* given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
* strided batched, but we can easily extend to other layouts. The returned offset can be either \p
* index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
* limitations.
*
* \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
* returns the 2D index of the tile that it computes. \see
* GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
*
* \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
* tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
* descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
* device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
* DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
* pointer offset into \p ComputePtrOffsetOfStridedBatch.
*
* \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
* Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
* realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
*
*/
template <typename GridwiseGemm,
typename ABDataType,
typename DsPointer,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename ComputePtrOffsetOfBatch,
typename Block2ETileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_batched_gemm_xdl(const ABDataType* __restrict__ p_a_grid,
const ABDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const index_t batch_count,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock_,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
DsPointer p_ds_grid_grp;
static constexpr index_t NumDTensor =
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_ds_grid_grp,
p_e_grid + e_batch_offset,
p_shared,
a_element_op,
b_element_op,
cde_element_op,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock_,
block_2_etile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = batch_count;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = compute_ptr_offset_of_batch;
ignore = block_2_etile_map;
#endif
}
template <typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
using DeviceOp = DeviceBatchedGemmMultiD_Xdl;
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
}();
return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
}
static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
}();
return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
}
template <typename ELay>
static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
{
const auto e_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideE, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideE));
}
}();
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
const std::array<index_t, NumDTensor>& NRaws,
const std::array<index_t, NumDTensor>& DsStride)
{
return generate_tuple(
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
},
Number<NumDTensor>{});
}
using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
struct ComputePtrOffsetOfStridedBatch
{
ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
index_t BatchStrideB,
std::array<ck::index_t, NumDTensor> BatchStrideDs,
index_t BatchStrideE)
: BatchStrideA_(BatchStrideA),
BatchStrideB_(BatchStrideB),
BatchStrideDs_(BatchStrideDs),
BatchStrideE_(BatchStrideE)
{
}
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_);
}
__host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
{
std::array<long_index_t, NumDTensor> ds_offset;
static_for<0, NumDTensor, 1>{}([&](auto i) {
ds_offset[i] = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]);
});
return ds_offset;
}
__host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideE_);
}
private:
index_t BatchStrideA_;
index_t BatchStrideB_;
std::array<ck::index_t, NumDTensor> BatchStrideDs_;
index_t BatchStrideE_;
};
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_M_K,
BGridDesc_N_K,
DsGridDesc_M_N,
EGridDesc_M_N,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEBlockTransferScalarPerVector_NPerBlock,
LoopSched>;
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
// Argument
struct Argument : public BaseArgument
{
Argument(const void* p_a_grid,
const void* p_b_grid,
std::array<const void*, NumDTensor> p_ds_grid,
void* p_e_grid,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t Batch,
index_t StrideA,
index_t StrideB,
const std::array<ck::index_t, NumDTensor>& StrideDs,
index_t StrideE,
index_t BatchStrideA,
index_t BatchStrideB,
const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
index_t BatchStrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_ds_grid_{},
p_e_grid_{static_cast<EDataType*>(p_e_grid)},
Batch_(Batch),
a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
ds_grid_desc_m_n_{},
e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
a_grid_desc_ak0_m_ak1_{
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
b_grid_desc_bk0_n_bk1_{
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
e_grid_desc_mblock_mperblock_nblock_nperblock_{},
compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE},
block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
// populate pointer, desc for Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
// D pointer
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
// D desc
ds_grid_desc_m_n_(i) =
DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaw, NRaw, StrideDs[i]);
});
// populate desc for Ds/E
if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
b_grid_desc_n_k_,
ds_grid_desc_m_n_,
e_grid_desc_m_n_,
block_2_etile_map_))
{
ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n_);
e_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
}
}
void Print() const
{
std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
static_for<0, NumDTensor, 1>{}(
[&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
}
// private:
// pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseGemm::DsGridPointer p_ds_grid_;
EDataType* p_e_grid_;
// Batch
index_t Batch_;
// tensor descriptors for problem definiton
AGridDesc_M_K a_grid_desc_m_k_;
BGridDesc_N_K b_grid_desc_n_k_;
DsGridDesc_M_N ds_grid_desc_m_n_;
EGridDesc_M_N e_grid_desc_m_n_;
// tensor descriptors for block/thread-wise copy
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock_;
// for calculating batch offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
// block-to-e-tile map
Block2ETileMap block_2_etile_map_;
// element-wise op
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceBatchedGemmMultiD_Xdl::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
arg.b_grid_desc_n_k_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_etile_map_))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
const index_t grid_size =
arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.Batch_;
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
auto launch_kernel = [&](auto has_main_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
const auto kernel = kernel_batched_gemm_xdl<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
typename GridwiseGemm::DsGridPointer,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
ComputePtrOffsetOfStridedBatch,
Block2ETileMap,
has_main_loop>;
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
arg.Batch_,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.compute_ptr_offset_of_batch_,
arg.block_2_etile_map_);
};
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
return launch_kernel(integral_constant<bool, true>{});
}
else
{
return launch_kernel(integral_constant<bool, false>{});
}
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
arg.b_grid_desc_n_k_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_etile_map_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t Batch,
index_t StrideA,
index_t StrideB,
const std::array<index_t, NumDTensor>& StrideDs,
index_t StrideE,
index_t BatchStrideA,
index_t BatchStrideB,
const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
index_t BatchStrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{p_a,
p_b,
p_ds,
p_e,
M,
N,
K,
Batch,
StrideA,
StrideB,
StrideDs,
StrideE,
BatchStrideA,
BatchStrideB,
BatchStrideDs,
BatchStrideE,
a_element_op,
b_element_op,
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t Batch,
index_t StrideA,
index_t StrideB,
const std::array<ck::index_t, NumDTensor>& StrideDs,
index_t StrideE,
index_t BatchStrideA,
index_t BatchStrideB,
const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
index_t BatchStrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) override
{
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
M,
N,
K,
Batch,
StrideA,
StrideB,
StrideDs,
StrideE,
BatchStrideA,
BatchStrideB,
BatchStrideDs,
BatchStrideE,
a_element_op,
b_element_op,
cde_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceBatchedGemmMultiD_Xdl"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock
<< AK1 << ", "
<< BK1 << ", "
<< getGemmSpecializationString(GemmSpec)
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm.hpp
View file @ 8a370fbb
......@@ -12,12 +12,6 @@ namespace ck {
namespace tensor_operation {
namespace device {
struct GemmShape
{
ck::index_t M, N, K;
ck::index_t StrideA, StrideB, StrideC;
};
template <typename ALayout,
typename BLayout,
typename CLayout,
......@@ -46,29 +40,6 @@ struct DeviceGemm : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceGroupedGemm : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(std::vector<const void*>& p_a,
std::vector<const void*>& p_b,
std::vector<void*>& p_c,
std::vector<GemmShape>& gemm_shapes,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGroupedGemmPtr = std::unique_ptr<
DeviceGroupedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_xdl_cshuffle.hpp
View file @ 8a370fbb
......@@ -349,13 +349,13 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
e_grid_desc_m_n_,
block_2_etile_map_))
{
e_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n_);
e_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
}
}
......@@ -411,8 +411,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
arg.e_grid_desc_m_n_,
arg.block_2_etile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
const index_t grid_size =
......
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