Skip to content

GitLab

Commit ed528d76 authored by Aleksander Dudek's avatar Aleksander Dudek
Browse files

Merge branch 'develop' into ck_tile_gemmkernel_reuse

parents 6e078dc0 355893cd
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1331 additions and 1106 deletions
Dockerfile
View file @ ed528d76
FROM ubuntu:20.04
ARG DEBIAN_FRONTEND=noninteractive
ARG ROCMVERSION=6.2
ARG ROCMVERSION=6.3
ARG compiler_version=""
ARG compiler_commit=""
ARG CK_SCCACHE=""
......@@ -13,17 +13,12 @@ RUN set -xe && \
apt-get update && apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl && \
curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN if [ "$ROCMVERSION" != "6.3" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.2.60200-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.2.60200-1_all.deb && \
RUN if [ "$ROCMVERSION" != "6.4" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.3.60300-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.3.60300-1_all.deb && \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "6.3" ] && [ "$compiler_version" = "rc1" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.3-20.04-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.3-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.3 rel-20 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2074281; \
fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" && \
......
Dockerfile.compiler
View file @ ed528d76
ARG BASE_DOCKER="rocm/composable_kernel:ck_ub20.04_rocm6.2"
ARG BASE_DOCKER="rocm/composable_kernel:ck_ub20.04_rocm6.3"
FROM $BASE_DOCKER
ARG compiler_version=""
ARG compiler_commit=""
......
Jenkinsfile
View file @ ed528d76
......@@ -38,7 +38,8 @@ def getBaseDockerImageName(){
img = "${params.USE_CUSTOM_DOCKER}"
}
else{
if (params.ROCMVERSION != "6.3"){
def ROCM_numeric = "${params.ROCMVERSION}" as float
if ( ROCM_numeric < 6.4 ){
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
else{
......@@ -329,10 +330,8 @@ def cmake_build(Map conf=[:]){
try{
archiveArtifacts "perf_fmha_fwd_*.log"
archiveArtifacts "perf_fmha_bwd_*.log"
stash name: "perf_fmha_fwd_gfx942.log"
stash name: "perf_fmha_bwd_gfx942.log"
stash name: "perf_fmha_fwd_gfx90a.log"
stash name: "perf_fmha_bwd_gfx90a.log"
stash includes: "perf_fmha_**_gfx942.log", name: "perf_fmha_log_gfx942"
stash includes: "perf_fmha_**_gfx90a.log", name: "perf_fmha_log_gfx90a"
}
catch(Exception err){
echo "could not locate the requested artifacts: ${err.getMessage()}. will skip the stashing."
......@@ -407,128 +406,6 @@ def buildHipClangJobAndReboot(Map conf=[:]){
}
}
def runCKProfiler(Map conf=[:]){
show_node_info()
env.HSA_ENABLE_SDMA=0
checkout scm
def image = getDockerImageName()
def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3')
def render_id = sh(returnStdout: true, script: 'getent group render | cut -d: -f3')
dockerOpts = dockerOpts + " --group-add=${video_id} --group-add=${render_id} "
echo "Docker flags: ${dockerOpts}"
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
def variant = env.STAGE_NAME
def retimage
gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
try {
(retimage, image) = getDockerImage(conf)
withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){
sh 'rocminfo | tee rocminfo.log'
if ( !runShell('grep -n "gfx" rocminfo.log') ){
throw new Exception ("GPU not found")
}
else{
echo "GPU is OK"
}
}
}
}
catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
echo "The job was cancelled or aborted"
throw e
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 24, unit: 'HOURS')
{
sh """
rm -rf build
mkdir build
"""
dir("build"){
unstash 'ckProfiler.tar.gz'
sh 'tar -xvf ckProfiler.tar.gz'
}
dir("script"){
if (params.RUN_FULL_QA){
sh "./run_full_performance_tests.sh 0 QA_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
archiveArtifacts "perf_batched_gemm.log"
archiveArtifacts "perf_grouped_gemm.log"
archiveArtifacts "perf_grouped_conv_fwd.log"
archiveArtifacts "perf_grouped_conv_bwd_data.log"
archiveArtifacts "perf_grouped_conv_bwd_weight.log"
archiveArtifacts "perf_gemm_bilinear.log"
archiveArtifacts "perf_reduction.log"
archiveArtifacts "perf_splitK_gemm.log"
archiveArtifacts "perf_onnx_gemm.log"
archiveArtifacts "perf_mixed_gemm.log"
// stash perf files to master
stash name: "perf_gemm.log"
stash name: "perf_resnet50_N256.log"
stash name: "perf_resnet50_N4.log"
stash name: "perf_batched_gemm.log"
stash name: "perf_grouped_gemm.log"
stash name: "perf_grouped_conv_fwd.log"
stash name: "perf_grouped_conv_bwd_data.log"
stash name: "perf_grouped_conv_bwd_weight.log"
stash name: "perf_gemm_bilinear.log"
stash name: "perf_reduction.log"
stash name: "perf_splitK_gemm.log"
stash name: "perf_onnx_gemm.log"
stash name: "perf_mixed_gemm.log"
//we will process results on the master node
}
else{
sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
// stash perf files to master
stash name: "perf_gemm.log"
stash name: "perf_resnet50_N256.log"
stash name: "perf_resnet50_N4.log"
//we will process the results on the master node
}
}
}
}
}
return retimage
}
def runPerfTest(Map conf=[:]){
try{
runCKProfiler(conf)
}
catch(e){
echo "throwing error exception in performance tests"
echo 'Exception occurred: ' + e.toString()
throw e
}
finally{
if (!conf.get("no_reboot", false)) {
reboot()
}
}
}
def Build_CK(Map conf=[:]){
show_node_info()
......@@ -588,36 +465,95 @@ def Build_CK(Map conf=[:]){
throw e
}
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 24, unit: 'HOURS')
timeout(time: 12, unit: 'HOURS')
{
//check whether to run performance tests on this node
def do_perf_tests = 0
def arch_type = 0
sh 'rocminfo | tee rocminfo.log'
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx1201" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
do_perf_tests = 1
echo "Stash profiler and run performance tests"
if ( runShell('grep -n "gfx90a" rocminfo.log') ){
arch_type = 1
}
else if ( runShell('grep -n "gfx942" rocminfo.log') ) {
arch_type = 2
}
else if ( runShell('grep -n "gfx1030" rocminfo.log') ) {
arch_type = 3
}
else if ( runShell('grep -n "gfx1101" rocminfo.log') ) {
arch_type = 4
}
else if ( runShell('grep -n "gfx1201" rocminfo.log') ) {
arch_type = 5
}
cmake_build(conf)
dir("build"){
//run tests and examples
//sh 'make -j check'
if (params.RUN_PERFORMANCE_TESTS && do_perf_tests == 0 ){
//we only need the ckProfiler to run the performance tests, so we pack and stash it
//do not stash profiler on nodes where we don't need to run performance tests
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash name: "ckProfiler.tar.gz"
}
if (params.RUN_FULL_QA && do_perf_tests == 0 ){
// build deb packages for all gfx9 targets and prepare to export
if (params.RUN_FULL_QA && arch_type == 1 ){
// build deb packages for all gfx9 targets on gfx90a system and prepare to export
echo "Build ckProfiler package"
sh 'make -j package'
archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb'
archiveArtifacts artifacts: 'composablekernel-tests_*.deb'
sh 'mv composablekernel-ckprofiler_*.deb ckprofiler_0.2.0_amd64.deb'
stash name: "ckprofiler_0.2.0_amd64.deb"
stash includes: "ckprofiler_0.2.0_amd64.deb", name: "ckprofiler_0.2.0_amd64.deb"
}
}
if (params.hipTensor_test && do_perf_tests == 0 ){
//build and test hipTensor
// run performance tests, stash the logs, results will be processed on the master node
dir("script"){
if (params.RUN_PERFORMANCE_TESTS){
if (params.RUN_FULL_QA && arch_type == 1){
// run full tests on gfx90a
echo "Run full performance tests"
sh "./run_full_performance_tests.sh 0 QA_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
archiveArtifacts "perf_batched_gemm.log"
archiveArtifacts "perf_grouped_gemm.log"
archiveArtifacts "perf_grouped_conv_fwd.log"
archiveArtifacts "perf_grouped_conv_bwd_data.log"
archiveArtifacts "perf_grouped_conv_bwd_weight.log"
archiveArtifacts "perf_gemm_bilinear.log"
archiveArtifacts "perf_reduction.log"
archiveArtifacts "perf_splitK_gemm.log"
archiveArtifacts "perf_onnx_gemm.log"
archiveArtifacts "perf_mixed_gemm.log"
stash includes: "perf_**.log", name: "perf_log"
}
else if ( arch_type == 1 ){
// run standard tests on gfx90a
echo "Run performance tests"
sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
archiveArtifacts "perf_gemm.log"
archiveArtifacts "perf_onnx_gemm.log"
archiveArtifacts "perf_resnet50_N256.log"
archiveArtifacts "perf_resnet50_N4.log"
stash includes: "perf_**.log", name: "perf_log"
}
// disable performance tests on gfx1030 for now.
//else if ( arch_type == 3){
// run basic tests on gfx1030
// echo "Run gemm performance tests"
// sh "./run_gemm_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME} gfx10"
// archiveArtifacts "perf_onnx_gemm_gfx10.log"
// stash includes: "perf_onnx_gemm_gfx10.log", name: "perf_log_gfx10"
//}
else if ( arch_type == 4){
// run basic tests on gfx11
echo "Run gemm performance tests"
sh "./run_gemm_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME} gfx11"
archiveArtifacts "perf_onnx_gemm_gfx11.log"
stash includes: "perf_onnx_gemm_gfx11.log", name: "perf_log_gfx11"
}
else if ( arch_type == 5 ){
// run basic tests on gfx12
echo "Run gemm performance tests"
sh "./run_gemm_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME} gfx12"
archiveArtifacts "perf_onnx_gemm_gfx12.log"
stash includes: "perf_onnx_gemm_gfx12.log", name: "perf_log_gfx12"
}
}
}
if (params.hipTensor_test && arch_type == 1 ){
// build and test hipTensor on gfx90a node
sh """#!/bin/bash
rm -rf "${params.hipTensor_branch}".zip
rm -rf hipTensor-"${params.hipTensor_branch}"
......@@ -689,10 +625,8 @@ def process_results(Map conf=[:]){
dir("script"){
if (params.RUN_CK_TILE_FMHA_TESTS){
try{
unstash "perf_fmha_fwd_gfx942.log"
unstash "perf_fmha_bwd_gfx942.log"
unstash "perf_fmha_fwd_gfx90a.log"
unstash "perf_fmha_bwd_gfx90a.log"
unstash "perf_fmha_log_gfx942"
unstash "perf_fmha_log_gfx90a"
}
catch(Exception err){
echo "could not locate the FMHA performance logs: ${err.getMessage()}."
......@@ -702,26 +636,26 @@ def process_results(Map conf=[:]){
// unstash perf files to master
unstash "ckprofiler_0.2.0_amd64.deb"
sh "sshpass -p ${env.ck_deb_pw} scp -o StrictHostKeyChecking=no ckprofiler_0.2.0_amd64.deb ${env.ck_deb_user}@${env.ck_deb_ip}:/var/www/html/composable_kernel/"
unstash "perf_gemm.log"
unstash "perf_resnet50_N256.log"
unstash "perf_resnet50_N4.log"
unstash "perf_batched_gemm.log"
unstash "perf_grouped_gemm.log"
unstash "perf_grouped_conv_fwd.log"
unstash "perf_grouped_conv_bwd_data.log"
unstash "perf_grouped_conv_bwd_weight.log"
unstash "perf_gemm_bilinear.log"
unstash "perf_reduction.log"
unstash "perf_splitK_gemm.log"
unstash "perf_onnx_gemm.log"
unstash "perf_mixed_gemm.log"
unstash "perf_log"
try{
unstash "perf_log_gfx11"
unstash "perf_log_gfx12"
}
catch(Exception err){
echo "could not locate the GEMM gfx11/gfx12 performance logs: ${err.getMessage()}."
}
sh "./process_qa_data.sh"
}
else{
// unstash perf files to master
unstash "perf_gemm.log"
unstash "perf_resnet50_N256.log"
unstash "perf_resnet50_N4.log"
unstash "perf_log"
try{
unstash "perf_log_gfx11"
unstash "perf_log_gfx12"
}
catch(Exception err){
echo "could not locate the GEMM gfx11/gfx12 performance logs: ${err.getMessage()}."
}
sh "./process_perf_data.sh"
}
}
......@@ -739,8 +673,8 @@ def process_results(Map conf=[:]){
}
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true;RUN_CODEGEN_TESTS=true
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.3;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.3;hipTensor_test=true;RUN_CODEGEN_TESTS=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
......@@ -765,8 +699,8 @@ pipeline {
description: 'If you want to use a custom docker image, please specify it here (default: leave blank).')
string(
name: 'ROCMVERSION',
defaultValue: '6.2',
description: 'Specify which ROCM version to use: 6.2 (default).')
defaultValue: '6.3',
description: 'Specify which ROCM version to use: 6.3 (default).')
string(
name: 'COMPILER_VERSION',
defaultValue: '',
......@@ -1240,29 +1174,6 @@ pipeline {
}
}
}
stage("Performance Tests")
{
parallel
{
stage("Run ckProfiler: gfx90a")
{
when {
beforeAgent true
expression { params.RUN_PERFORMANCE_TESTS.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
}
options { retry(1) }
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "NO_CK_BUILD"
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
cleanWs()
}
}
}
}
stage("Process Performance Test Results")
{
parallel
......
example/62_convnd_activ/convscale/convnd_fwd_convscale_common.hpp
View file @ ed528d76
......@@ -172,12 +172,13 @@ bool run_grouped_conv_fwd(bool do_verification,
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
// values generated: -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 6});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ ed528d76
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -106,8 +106,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
static constexpr auto I3 = Number<3>{};
static constexpr index_t KPerBlock = K0PerBlock * K1;
static constexpr auto transform_conv_to_gemm =
TransformConvBwdDataToGemm_v1<NDimSpatial,
using ConvToGemmBwdDataTransform = TransformConvBwdDataToGemm_v1<NDimSpatial,
ConvBackwardDataSpecialization,
K1,
K1,
......@@ -115,80 +114,27 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
NPerBlock,
KPerBlock,
true /* DoPadGemmM */,
true /* DoPadGemmN */>{};
true /* DoPadGemmN */,
ALayout,
BLayout,
ELayout>;
static auto GetDummyABDsEGridDescriptor()
static auto
GetDummyABDsEGridDescriptor(const ConvToGemmBwdDataTransform& conv_to_gemm_transform)
{
const std::array<index_t, NDimSpatial + 3> dummy_tensor_lengths = {1};
const std::array<index_t, NDimSpatial + 3> dummy_tensor_strides = {1};
const std::array<index_t, NDimSpatial> dummy_spatial_lengths = {1};
const auto a_grid_desc_ak0_m_ak1 =
transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
const auto b_grid_desc_bk0_n_bk1 =
transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
const auto ds_grid_desc_m_n = generate_tuple(
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
return transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
},
const auto a_grid_desc_ak0_m_ak1 = conv_to_gemm_transform.MakeADescriptor_AK0_M_AK1();
const auto b_grid_desc_bk0_n_bk1 = conv_to_gemm_transform.MakeBDescriptor_BK0_N_BK1();
const auto ds_grid_desc_m_n =
generate_tuple([&](auto) { return conv_to_gemm_transform.MakeCDescriptor_M_N(); },
Number<NumDTensor>{});
const auto e_grid_desc_m_n =
transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
const auto e_grid_desc_m_n = conv_to_gemm_transform.MakeCDescriptor_M_N();
return make_tuple(
a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, ds_grid_desc_m_n, e_grid_desc_m_n);
}
// desc
using ABDsEGridDesc = decltype(GetDummyABDsEGridDescriptor());
constexpr static ConvToGemmBwdDataTransform dummy_conv_to_gemm_transform;
using ABDsEGridDesc = decltype(GetDummyABDsEGridDescriptor(dummy_conv_to_gemm_transform));
using AGridDesc_AK0_M_AK1 = remove_cvref_t<tuple_element_t<0, ABDsEGridDesc>>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<tuple_element_t<1, ABDsEGridDesc>>;
......@@ -270,7 +216,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_c_wis_lengths,
/*ds_g_n_c_wis_lengths*/,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths,
......@@ -291,15 +237,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
b_element_op_{b_element_op},
cde_element_op_{cde_element_op},
a_g_n_k_wos_lengths_{a_g_n_k_wos_lengths},
a_g_n_k_wos_strides_{a_g_n_k_wos_strides},
b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
ds_g_n_c_wis_lengths_{ds_g_n_c_wis_lengths},
ds_g_n_c_wis_strides_{ds_g_n_c_wis_strides},
e_g_n_c_wis_lengths_{e_g_n_c_wis_lengths},
e_g_n_c_wis_strides_{e_g_n_c_wis_strides},
conv_filter_strides_{conv_filter_strides},
conv_filter_dilations_{conv_filter_dilations},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
......@@ -382,9 +321,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
tildes = {i_ztilde, i_ytilde, i_xtilde};
}
const auto a_grid_desc_ak0_m_ak1 =
transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
a_g_n_k_wos_lengths,
ConvToGemmBwdDataTransform conv_to_gemm_transform_{a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
......@@ -394,56 +331,37 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
conv_filter_dilations,
input_left_pads,
input_right_pads,
tildes);
tildes};
const auto a_grid_desc_ak0_m_ak1 =
conv_to_gemm_transform_.MakeADescriptor_AK0_M_AK1();
const auto b_grid_desc_bk0_n_bk1 =
transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_c_wis_lengths,
e_g_n_c_wis_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
tildes);
conv_to_gemm_transform_.MakeBDescriptor_BK0_N_BK1();
DsGridDesc_M_N ds_grid_desc_m_n;
// populate Ds desc
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
ds_grid_desc_m_n(i) =
transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
static_assert(is_same_v<DLayout, ELayout>);
ConvToGemmBwdDataTransform conv_to_gemm_transform_d{
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
ds_g_n_c_wis_lengths[i],
e_g_n_c_wis_lengths,
ds_g_n_c_wis_strides[i],
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
tildes);
tildes};
ds_grid_desc_m_n(i) = conv_to_gemm_transform_d.MakeCDescriptor_M_N();
});
const auto e_grid_desc_m_n =
transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_c_wis_lengths,
e_g_n_c_wis_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
tildes);
const auto e_grid_desc_m_n = conv_to_gemm_transform_.MakeCDescriptor_M_N();
// for check validity
ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n);
......@@ -522,17 +440,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
BElementwiseOp b_element_op_;
CDEElementwiseOp cde_element_op_;
// for checking IsSupportedArgument()
std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_lengths_;
std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_strides_;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_lengths_;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_strides_;
std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_lengths_;
std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_strides_;
std::array<index_t, NDimSpatial> conv_filter_strides_;
std::array<index_t, NDimSpatial> conv_filter_dilations_;
std::array<index_t, NDimSpatial> input_left_pads_;
std::array<index_t, NDimSpatial> input_right_pads_;
};
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp
View file @ ed528d76
......@@ -998,14 +998,14 @@ struct FmhaFwdKernel
return pad_tensor_view(
q_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kSubQKHeaddim>{}),
sequence<false, kPadHeadDimQ>{});
sequence<kPadSeqLenQ, kPadHeadDimQ>{});
}
else
{
return pad_tensor_view(
q_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{}),
sequence<false, kPadHeadDimQ>{});
sequence<kPadSeqLenQ, kPadHeadDimQ>{});
}
}();
const auto k_dram = [&]() {
......@@ -1019,7 +1019,7 @@ struct FmhaFwdKernel
return pad_tensor_view(
k_dram_naive,
make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}),
sequence<false, kPadHeadDimQ>{});
sequence<kPadSeqLenK, kPadHeadDimQ>{});
}();
const auto v_dram = [&]() {
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
......@@ -1041,7 +1041,7 @@ struct FmhaFwdKernel
return pad_tensor_view(
v_dram_transposed,
make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
sequence<kPadHeadDimV, false>{});
sequence<kPadHeadDimV, kPadSeqLenK>{});
}
else
{
......@@ -1055,7 +1055,7 @@ struct FmhaFwdKernel
return pad_tensor_view(
v_dram_naive,
make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
sequence<false, kPadSeqLenK>{});
sequence<kPadHeadDimV, kPadSeqLenK>{});
}
}();
......@@ -1097,8 +1097,9 @@ struct FmhaFwdKernel
number<FmhaPipeline::kAlignmentBias>{},
number<1>{});
return pad_tensor_view(
bias_dram_naive, bias_dram_window_lengths, sequence<false, kPadSeqLenK>{});
return pad_tensor_view(bias_dram_naive,
bias_dram_window_lengths,
sequence<kPadSeqLenQ, kPadSeqLenK>{});
}();
return make_tile_window(bias_dram, bias_dram_window_lengths, {i_m0, 0});
......
library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt
View file @ ed528d76
......@@ -6,7 +6,7 @@ set(CK_TILE_SRC_FOLDER ${CMAKE_SOURCE_DIR}/include/ck_tile/)
# CK Codegen requires dataclass which is added in Python 3.7
# Python version 3.8 is required for general good practice as it is default for Ubuntu 20.04
if(NOT CK_USE_ALTERNATIVE_PYTHON)
find_package(PythonInterp 3 REQUIRED)
find_package(Python3 COMPONENTS Interpreter Development)
else()
message("Using alternative python version")
set(EXTRA_PYTHON_PATH)
......@@ -33,7 +33,7 @@ set(FMHA_KNOWN_APIS "fwd,fwd_splitkv,fwd_appendkv,bwd")
# Note: The receipt 3 arg filters the generated backwards instances to reduce compilation time.
# With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
execute_process(
COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
COMMAND ${Python3_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
--list_blobs ${FMHA_CPP_FOLDER}/blob_list.txt
--api ${FMHA_KNOWN_APIS}
--receipt 3
......@@ -50,7 +50,7 @@ endif()
# With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
add_custom_command(
OUTPUT ${FMHA_GEN_BLOBS}
COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
COMMAND ${Python3_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
--output_dir ${FMHA_CPP_FOLDER}
--api ${FMHA_KNOWN_APIS}
--receipt 3
......
script/process_perf_data.py
View file @ ed528d76
......@@ -82,7 +82,7 @@ def parse_logfile(logfile):
StrideA=[]
StrideB=[]
StrideC=[]
if 'perf_gemm.log' in logfile:
if 'perf_gemm' in logfile and 'gemm_bilinear' not in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
......@@ -260,7 +260,7 @@ def main():
conn = sqlEngine.connect()
#save gemm performance tests:
if 'perf_gemm.log' in filename:
if 'perf_gemm' in filename and 'gemm_bilinear' not in filename:
#write the ck_gemm_test_params table only needed once the test set changes
#post_test_params(test_list,conn)
for i in range(1,len(results)+1):
......
script/process_perf_data.sh
View file @ ed528d76
......@@ -11,9 +11,22 @@
#process results
python3 process_perf_data.py perf_gemm.log
python3 process_perf_data.py perf_onnx_gemm.log
python3 process_perf_data.py perf_resnet50_N256.log
python3 process_perf_data.py perf_resnet50_N4.log
file=./perf_onnx_gemm_gfx10.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx10.log
fi
file=./perf_onnx_gemm_gfx11.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx11.log
fi
file=./perf_onnx_gemm_gfx12.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx12.log
fi
file=./perf_fmha_fwd_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_fmha_fwd_gfx942.log
......
script/process_qa_data.sh
View file @ ed528d76
......@@ -24,6 +24,18 @@ python3 process_perf_data.py perf_splitK_gemm.log
python3 process_perf_data.py perf_onnx_gemm.log
python3 process_perf_data.py perf_mixed_gemm.log
file=./perf_onnx_gemm_gfx10.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx10.log
fi
file=./perf_onnx_gemm_gfx11.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx11.log
fi
file=./perf_onnx_gemm_gfx12.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx12.log
fi
file=./perf_fmha_fwd_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_fmha_fwd_gfx942.log
......
script/run_full_performance_tests.sh
View file @ ed528d76
......@@ -5,7 +5,7 @@
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <branch name> <node name>
# input arguments:
# verification = 0 : do not verify result correctness on CPU
# = 1 : verifuy correctness on CPU (may take a long time)
......
script/run_gemm_performance_tests.sh 0 → 100755
View file @ ed528d76
#!/bin/bash
#
# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
# run the script as "./run_gemm_performance_tests.sh <verification> <tag for your test environment> <branch name> <node name> <arch>
# input arguments:
# verification = 0 : do not verify result correctness on CPU
# = 1 : verify correctness on CPU (may take a long time)
# environment tag : a string describing the specifics of your test environment
# branch name : name of the branch in git repo (git status | grep -e 'On branch')
# node name : $hostname
# arch : GPU architecture, e.g. "gfx9" or "gfx1100"
#get the command line arguments:
export verify=$1
echo 'Verification: ' $verify
export env_type=$2
echo 'Environment type: ' $env_type
export branch=$3
echo 'Branch name: ' $branch
export host_name=$4
echo 'Host name: ' $host_name
export arch=$5
echo 'GPU architecture: ' $arch
function print_log_header(){
rm -f $1;
echo 'On branch ' $3 &> $1;
echo 'Node name: ' $4 >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >> $1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
#run ONNX gemm tests
export onnx_log="perf_onnx_gemm_$arch.log"
print_log_header $onnx_log $env_type $branch $host_name
./profile_onnx_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
./profile_onnx_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
script/run_performance_tests.sh
View file @ ed528d76
#!/bin/bash
#
# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
# run the script as "./run_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
# run the script as "./run_performance_tests.sh <verification> <tag for your test environment> <branch name> <node name>
# input arguments:
# verification = 0 : do not verify result correctness on CPU
# = 1 : verify correctness on CPU (may take a long time)
......@@ -51,20 +51,11 @@ print_log_header $gemm_log $env_type $branch $host_name
./profile_gemm.sh gemm 2 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 $verify 1 0 1 | tee -a $gemm_log
#run grouped_fwd fp16 tests
export grouped_conv_fwd_log="perf_grouped_conv_fwd_fp16.log"
print_log_header $conv_fwd_log $env_type $branch $host_name
./profile_grouped_conv_fwd.sh grouped_conv_fwd 1 1 0 $verify 1 0 1 256 2>&1 | tee -a $grouped_conv_fwd_log
#run grouped_bwd_data fp16 tests
export grouped_conv_bwd_data_log="perf_grouped_conv_bwd_data_fp16.log"
print_log_header $grouped_conv_bwd_data_log $env_type $branch $host_name
./profile_grouped_conv_bwd_data.sh grouped_conv_bwd_data 1 1 $verify 1 0 1 256 2>&1 | tee -a $grouped_conv_bwd_data_log
#run grouped_bwd_weight fp16 tests
export grouped_conv_bwd_weight_log="perf_grouped_conv_bwd_weight_fp16.log"
print_log_header $grouped_conv_bwd_weight_log $env_type $branch $host_name
./profile_grouped_conv_bwd_weight.sh grouped_conv_bwd_weight 1 1 $verify 1 0 1 256 1 2>&1 | tee -a $grouped_conv_bwd_weight_log
#run ONNX gemm tests
export onnx_log="perf_onnx_gemm.log"
print_log_header $onnx_log $env_type $branch $host_name
./profile_onnx_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
./profile_onnx_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
#run resnet50 tests
export resnet256_log="perf_resnet50_N256.log"
......
test/data_type/test_custom_type.cpp
View file @ ed528d76
......@@ -51,8 +51,11 @@ TEST(Custom_bool, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bool_t>()(Number<i>{}) = custom_bool_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bool_t, size> left_vec{right_vec};
vector_type<custom_bool_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_bool_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bool_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -129,8 +132,11 @@ TEST(Custom_int8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_int8_t>()(Number<i>{}) = custom_int8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_int8_t, size> left_vec{right_vec};
vector_type<custom_int8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_int8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_int8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -207,8 +213,11 @@ TEST(Custom_uint8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_uint8_t>()(Number<i>{}) = custom_uint8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_uint8_t, size> left_vec{right_vec};
vector_type<custom_uint8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_uint8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -287,8 +296,11 @@ TEST(Custom_f8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_f8_t>()(Number<i>{}) = custom_f8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_f8_t, size> left_vec{right_vec};
vector_type<custom_f8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_f8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_f8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -369,8 +381,11 @@ TEST(Custom_bf8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bf8_t>()(Number<i>{}) = custom_bf8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bf8_t, size> left_vec{right_vec};
vector_type<custom_bf8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_bf8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -450,8 +465,11 @@ TEST(Custom_half, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_half_t>()(Number<i>{}) = custom_half_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_half_t, size> left_vec{right_vec};
vector_type<custom_half_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_half_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_half_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -533,8 +551,11 @@ TEST(Custom_bhalf, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bhalf_t>()(Number<i>{}) = custom_bhalf_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bhalf_t, size> left_vec{right_vec};
vector_type<custom_bhalf_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_bhalf_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -615,8 +636,11 @@ TEST(Custom_float, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_float_t>()(Number<i>{}) = custom_float_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_float_t, size> left_vec{right_vec};
vector_type<custom_float_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_float_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_float_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -693,8 +717,11 @@ TEST(Custom_double, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_double_t>()(Number<i>{}) = custom_double_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_double_t, size> left_vec{right_vec};
vector_type<custom_double_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_double_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_double_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -813,8 +840,11 @@ TEST(Complex_half, TestAsType)
right_vec.template AsType<complex_half_t>()(Number<i>{}) =
complex_half_t{test_vec.at(num_elem * i), test_vec.at(num_elem * i + 1)};
});
// copy the vector
vector_type<complex_half_t, size> left_vec{right_vec};
vector_type<complex_half_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<complex_half_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).real,
......@@ -907,8 +937,11 @@ TEST(FP8OCP, TestAsType)
right_vec.template AsType<f8_t>()(Number<i>{}) = ck::type_convert<f8_t>(test_vec.at(i));
});
// copy the vector
vector_type<f8_t, size> left_vec{right_vec};
vector_type<f8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<f8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
......@@ -984,8 +1017,11 @@ TEST(BF8OCP, TestAsType)
right_vec.template AsType<bf8_t>()(Number<i>{}) = ck::type_convert<bf8_t>(test_vec.at(i));
});
// copy the vector
vector_type<bf8_t, size> left_vec{right_vec};
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<bf8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
......
test/grouped_convnd_bwd_data/CMakeLists.txt
View file @ ed528d76
add_gtest_executable(test_grouped_convnd_bwd_data test_grouped_convnd_bwd_data_xdl_wmma.cpp)
add_gtest_executable(test_grouped_convnd_bwd_data_xdl test_grouped_convnd_bwd_data_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_data PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
target_link_libraries(test_grouped_convnd_bwd_data_xdl PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
endif()
add_gtest_executable(test_grouped_convnd_bwd_data_wmma test_grouped_convnd_bwd_data_wmma.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_data_wmma PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
endif()
add_gtest_executable(test_grouped_convnd_bwd_data_interface_xdl test_grouped_convnd_bwd_data_interface_xdl.cpp)
if(result EQUAL 0)
......
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_wmma.cpp 0 → 100644
View file @ ed528d76
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdDataWmma : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
using OutLayout = std::tuple_element_t<1, Tuple>;
using WeiLayout = std::tuple_element_t<2, Tuple>;
using InLayout = std::tuple_element_t<3, Tuple>;
std::vector<ck::utils::conv::ConvParam> conv_params;
template <ck::index_t NDimSpatial>
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto& param : conv_params)
{
pass = pass && ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
OutLayout,
WeiLayout,
InLayout,
DataType,
DataType,
DataType>(
true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param);
}
EXPECT_TRUE(pass);
}
};
using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<int8_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<int8_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<int8_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<int8_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdDataWmma2d : public TestGroupedConvndBwdDataWmma<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdDataWmma3d : public TestGroupedConvndBwdDataWmma<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdDataWmma2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataWmma3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdDataWmma2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdDataWmma3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->template Run<3>();
}
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_xdl_wmma.cpp → test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_xdl.cpp
View file @ ed528d76
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
......@@ -12,7 +12,7 @@
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdData : public ::testing::Test
class TestGroupedConvndBwdDataXdl : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
......@@ -51,35 +51,31 @@ using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>,
std::tuple<int8_t, GNHWK, GKYXC, GNHWC>,
std::tuple<float, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>,
std::tuple<int8_t, NHWGK, GKYXC, NHWGC>>;
std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::bhalf_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<int8_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<float, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<int8_t, NDHWGK, GKZYXC, NDHWGC>>;
std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdData2d : public TestGroupedConvndBwdData<Tuple>
class TestGroupedConvndBwdDataXdl2d : public TestGroupedConvndBwdDataXdl<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdData3d : public TestGroupedConvndBwdData<Tuple>
class TestGroupedConvndBwdDataXdl3d : public TestGroupedConvndBwdDataXdl<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdData2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdData3d, KernelTypes3d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataXdl2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataXdl3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
TYPED_TEST(TestGroupedConvndBwdDataXdl2d, Test2D)
{
this->conv_params.clear();
......@@ -94,10 +90,13 @@ TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
// SplitN case
this->conv_params.push_back(
{2, 1, 128, 4, 192, {2, 2}, {224, 224}, {224, 224}, {1, 1}, {0, 0}, {0, 0}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
TYPED_TEST(TestGroupedConvndBwdDataXdl3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
......@@ -112,5 +111,17 @@ TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
// SplitN case
this->conv_params.push_back({3,
1,
128,
4,
192,
{2, 2, 2},
{2, 224, 224},
{1, 224, 224},
{1, 1, 1},
{0, 0, 0},
{0, 0, 0}});
this->template Run<3>();
}
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