Skip to content

GitLab

Commit 86f8ac01 authored by Jakub Piasecki's avatar Jakub Piasecki
Browse files

Merge remote-tracking branch 'origin/develop' into jakpias/pool1d_fwd

parents 3f6360d0 ab60b390
Hide whitespace changes
Inline Side-by-side
Showing with 195 additions and 205 deletions
CMakeLists.txt
View file @ 86f8ac01
......@@ -189,7 +189,9 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
message("Adding the enable-post-misched=0 compiler flag")
add_compile_options("SHELL: -mllvm -enable-post-misched=0")
endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000)
set(check-coerce)
check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce)
if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000)
message("Adding the amdgpu-coerce-illegal-types=1")
add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
endif()
......@@ -539,6 +541,9 @@ if(NOT DEFINED INSTANCES_ONLY)
PACKAGE_NAME examples
)
add_subdirectory(example)
if(GPU_TARGETS MATCHES "gfx9" AND NOT INSTANCES_ONLY)
add_subdirectory(codegen)
endif()
if(BUILD_TESTING)
add_subdirectory(test)
endif()
......
Dockerfile
View file @ 86f8ac01
FROM ubuntu:20.04
ARG DEBIAN_FRONTEND=noninteractive
ARG ROCMVERSION=6.1
ARG ROCMVERSION=6.2
ARG compiler_version=""
ARG compiler_commit=""
ARG CK_SCCACHE=""
......@@ -17,17 +17,12 @@ RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN if [ "$ROCMVERSION" != "6.2" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/6.1/ubuntu/focal/amdgpu-install_6.1.60100-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.1.60100-1_all.deb && \
RUN if [ "$ROCMVERSION" != "6.3" ]; then \
sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.2.60200-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.2.60200-1_all.deb && \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "6.2" ] && [ "$compiler_version" = "rc4" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.2-20.04-1_all.deb --no-check-certificate" && \
apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.2-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.2 rel-63 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=2009461; \
fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
......@@ -64,6 +59,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
python3-dev \
python3-pip \
redis \
rocm-llvm-dev \
sshpass \
stunnel \
software-properties-common \
......
Jenkinsfile
View file @ 86f8ac01
......@@ -38,7 +38,7 @@ def getDockerImageName(){
img = "${params.USE_CUSTOM_DOCKER}"
}
else{
if (params.ROCMVERSION != "6.2"){
if (params.ROCMVERSION != "6.3"){
if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
......@@ -204,6 +204,9 @@ def cmake_build(Map conf=[:]){
cd build
"""
def invocation_tag=""
if (setup_args.contains("gfx12")){
invocation_tag="gfx12"
}
if (setup_args.contains("gfx11")){
invocation_tag="gfx11"
}
......@@ -285,6 +288,19 @@ def cmake_build(Map conf=[:]){
if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) {
archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true
}
if (params.RUN_CK_TILE_TESTS){
try{
archiveArtifacts "perf_fmha_fwd_*.log"
archiveArtifacts "perf_fmha_bwd_*.log"
stash name: "perf_fmha_fwd_gfx942.log"
stash name: "perf_fmha_bwd_gfx942.log"
stash name: "perf_fmha_fwd_gfx90a.log"
stash name: "perf_fmha_bwd_gfx90a.log"
}
catch(Exception err){
echo "could not locate the requested artifacts: ${err.getMessage()}. will skip the stashing."
}
}
}
def buildHipClangJob(Map conf=[:]){
......@@ -518,7 +534,7 @@ def Build_CK(Map conf=[:]){
//check whether to run performance tests on this node
def do_perf_tests = 0
sh 'rocminfo | tee rocminfo.log'
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx1201" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
do_perf_tests = 1
echo "Stash profiler and run performance tests"
}
......@@ -612,6 +628,17 @@ def process_results(Map conf=[:]){
timeout(time: 1, unit: 'HOURS'){
try{
dir("script"){
if (params.RUN_CK_TILE_TESTS){
try{
unstash "perf_fmha_fwd_gfx942.log"
unstash "perf_fmha_bwd_gfx942.log"
unstash "perf_fmha_fwd_gfx90a.log"
unstash "perf_fmha_bwd_gfx90a.log"
}
catch(Exception err){
echo "could not locate the FMHA performance logs: ${err.getMessage()}."
}
}
if (params.RUN_FULL_QA){
// unstash perf files to master
unstash "ckprofiler_0.2.0_amd64.deb"
......@@ -652,10 +679,10 @@ def process_results(Map conf=[:]){
}
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.1; RUN_CK_TILE_TESTS=true
0 21 * * * % ROCMVERSION=6.1;hipTensor_test=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2; RUN_CK_TILE_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
pipeline {
......@@ -677,8 +704,8 @@ pipeline {
description: 'If you want to use a custom docker image, please specify it here (default: leave blank).')
string(
name: 'ROCMVERSION',
defaultValue: '6.1',
description: 'Specify which ROCM version to use: 6.1 (default).')
defaultValue: '6.2',
description: 'Specify which ROCM version to use: 6.2 (default).')
string(
name: 'COMPILER_VERSION',
defaultValue: '',
......@@ -719,10 +746,6 @@ pipeline {
name: "RUN_PERFORMANCE_TESTS",
defaultValue: true,
description: "Run the performance tests (default: ON)")
booleanParam(
name: "RUN_CODEGEN_TESTS",
defaultValue: true,
description: "Run the codegen tests (default: ON)")
booleanParam(
name: "RUN_CK_TILE_TESTS",
defaultValue: false,
......@@ -731,6 +754,11 @@ pipeline {
name: "BUILD_INSTANCES_ONLY",
defaultValue: false,
description: "Test building instances for various architectures simultaneously (default: OFF)")
booleanParam(
name: "BUILD_GFX12",
defaultValue: false,
description: "Build CK and run tests on gfx12 (default: OFF)")
}
environment{
dbuser = "${dbuser}"
......@@ -809,33 +837,6 @@ pipeline {
}
}
}
stage("Run Codegen Tests")
{
parallel
{
stage("Run Codegen Tests on gfx90a")
{
when {
beforeAgent true
expression { params.RUN_CODEGEN_TESTS.toBoolean() }
}
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "NO_CK_BUILD"
execute_args = """ cd ../codegen && rm -rf build && mkdir build && cd build && \
cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ \
-D CMAKE_BUILD_TYPE=Release \
-D GPU_TARGETS="gfx90a" \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j check"""
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
cleanWs()
}
}
}
}
stage("Run CK_TILE Tests")
{
parallel
......@@ -852,8 +853,7 @@ pipeline {
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx90a && \
make -j64 tile_example_fmha_fwd tile_example_fmha_bwd && \
cd ../ &&
example/ck_tile/01_fmha/script/smoke_test_fwd.sh && \
example/ck_tile/01_fmha/script/smoke_test_bwd.sh"""
example/ck_tile/01_fmha/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx90a """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
......@@ -872,8 +872,7 @@ pipeline {
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx942 && \
make -j64 tile_example_fmha_fwd tile_example_fmha_bwd && \
cd ../ &&
example/ck_tile/01_fmha/script/smoke_test_fwd.sh && \
example/ck_tile/01_fmha/script/smoke_test_bwd.sh"""
example/ck_tile/01_fmha/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx942 """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
......@@ -1008,6 +1007,26 @@ pipeline {
cleanWs()
}
}
stage("Build CK and run Tests on gfx1201")
{
when {
beforeAgent true
expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
}
agent{ label rocmnode("gfx1201") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1201" -DDL_KERNELS=ON -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx1201" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
cleanWs()
}
}
}
}
......
codegen/CMakeLists.txt
View file @ 86f8ac01
cmake_minimum_required(VERSION 3.16)
project(composable_kernel_host LANGUAGES CXX HIP)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
......@@ -8,17 +5,9 @@ set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/..)
find_package(ROCM)
include(ROCMInstallTargets)
include(ROCMTest)
add_compile_options(-std=c++17)
find_package(hip)
## HIP
set(CMAKE_HIP_PLATFORM amd)
set(CMAKE_HIP_COMPILER ${CMAKE_CXX_COMPILER})
set(CMAKE_HIP_EXTENSIONS ON)
message("CMAKE_HIP_COMPILER: ${CMAKE_HIP_COMPILER}")
add_custom_target(codegen)
# add include directories
include_directories(BEFORE
......@@ -32,8 +21,9 @@ list(APPEND CMAKE_MODULE_PATH ${CK_ROOT}/cmake)
include(Embed)
file(GLOB_RECURSE KERNEL_FILES CONFIGURE_DEPENDS
${CK_ROOT}/include/ck/*.hpp)
message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
message(STATUS "RELATIVE: ${CK_ROOT}/include")
#printouts fot debug purposes
#message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
#message(STATUS "RELATIVE: ${CK_ROOT}/include")
add_embed_library(ck_headers ${KERNEL_FILES} RELATIVE ${CK_ROOT}/include)
file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
......
codegen/include/ck/host/types.hpp
View file @ 86f8ac01
......@@ -76,8 +76,11 @@ std::string SequenceStr(const std::vector<int>& v);
std::string MakeTuple(const std::vector<std::string>& v);
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wglobal-constructors"
template <int... xs>
const std::string S = SequenceStr({xs...});
#pragma clang diagnostic pop
constexpr const char* PassThrough = "ck::tensor_operation::element_wise::PassThrough";
constexpr const char* Bilinear = "ck::tensor_operation::element_wise::Bilinear";
......
codegen/src/device_gemm_multiple_d_operation_xdl_cshuffle.cpp
View file @ 86f8ac01
......@@ -3,6 +3,7 @@
#include "ck/host/device_gemm_multiple_d/operation.hpp"
#include "ck/host/stringutils.hpp"
#include "ck/host/types.hpp"
#include "ck/host/utils.hpp"
#include <cassert>
......@@ -32,11 +33,11 @@ static std::string GetGemmSpec(const std::size_t m,
}
// function to update prologue/epilogue with user provided operation
void Operation_Xdl_CShuffle::update_prologue(const std::string& prologue)
void Operation_Xdl_CShuffle::update_prologue(const std::string& pro)
{
if(!prologue.empty())
if(!pro.empty())
{
this->prologue = prologue;
this->prologue = pro;
this->cde_elem_op = "CDEElementOp";
}
else
......@@ -45,11 +46,11 @@ void Operation_Xdl_CShuffle::update_prologue(const std::string& prologue)
}
}
void Operation_Xdl_CShuffle::update_epilogue(const std::string& epilogue)
void Operation_Xdl_CShuffle::update_epilogue(const std::string& epi)
{
if(!epilogue.empty())
if(!epi.empty())
{
this->epilogue = epilogue;
this->epilogue = epi;
this->cde_elem_op = "CDEElementOp";
}
else
......
codegen/src/device_grouped_conv_fwd_multiple_abd_operation_xdl_cshuffle.cpp
View file @ 86f8ac01
......@@ -4,6 +4,7 @@
#include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
#include <iostream>
#include "ck/host/stringutils.hpp"
#include "ck/host/types.hpp"
#include "ck/host/utils.hpp"
#include <cassert>
......@@ -11,34 +12,15 @@ namespace ck {
namespace host {
namespace conv {
// calculate appropriate Gemm Specification based on input tensor dimensions
// NOTE: in CK, MNKPadding is always used for forward convolution
static std::string GetGemmSpec(const std::size_t m,
const std::size_t n,
const std::size_t k,
const std::size_t m_per_block,
const std::size_t n_per_block,
const std::size_t k_per_block)
{
std::string spec = "";
if(integer_divide_ceil(m, m_per_block) * m_per_block - m != 0)
spec += "M";
if(integer_divide_ceil(n, n_per_block) * n_per_block - n != 0)
spec += "N";
if(integer_divide_ceil(k, k_per_block) * k_per_block - k != 0)
spec += "K";
if(spec == "")
return "ck::tensor_operation::device::GemmSpecialization::Default";
return "ck::tensor_operation::device::GemmSpecialization::" + spec + "Padding";
}
// NOTE: in CK, MNKPadding is always used for forward convolution, so didn't
// add GemmSpec function here
// function to update prologue/epilogue with user provided operation
void Operation_Conv_Fwd_Xdl_Cshuffle::update_prologue(const std::string& prologue)
void Operation_Conv_Fwd_Xdl_Cshuffle::update_prologue(const std::string& pro)
{
if(!prologue.empty())
if(!pro.empty())
{
this->prologue = prologue;
this->prologue = pro;
this->cde_elem_op = "CDEElementOp";
}
else
......@@ -47,11 +29,11 @@ void Operation_Conv_Fwd_Xdl_Cshuffle::update_prologue(const std::string& prologu
}
}
void Operation_Conv_Fwd_Xdl_Cshuffle::update_epilogue(const std::string& epilogue)
void Operation_Conv_Fwd_Xdl_Cshuffle::update_epilogue(const std::string& epi)
{
if(!epilogue.empty())
if(!epi.empty())
{
this->epilogue = epilogue;
this->epilogue = epi;
this->cde_elem_op = "CDEElementOp";
}
else
......
codegen/src/headers.cpp
View file @ 86f8ac01
......@@ -4,7 +4,10 @@
namespace ck {
namespace host {
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wglobal-constructors"
const std::string config_header = "";
#pragma clang diagnostic pop
std::unordered_map<std::string_view, std::string_view> GetHeaders()
{
......
codegen/test/CMakeLists.txt
View file @ 86f8ac01
......@@ -4,7 +4,9 @@ file(GLOB TEST_SRCS CONFIGURE_DEPENDS *.cpp)
foreach(TEST_SRC ${TEST_SRCS})
set_source_files_properties(${TEST_SRC} PROPERTIES LANGUAGE HIP)
get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
rocm_add_test_executable(test_host_${BASE_NAME} ${TEST_SRC})
add_executable(test_host_${BASE_NAME} ${TEST_SRC})
add_dependencies(codegen test_host_${BASE_NAME})
add_test(NAME codegen_test_${BASE_NAME} COMMAND test_host_${BASE_NAME})
target_link_libraries(test_host_${BASE_NAME} ck_rtc ck_host)
# target_link_libraries(test_host_${BASE_NAME} ${CK_ROOT}/build/lib/libutility.a)
target_include_directories(test_host_${BASE_NAME} PUBLIC include())
......
codegen/test/grouped_conv_fwd_multiple_d_v1.cpp
View file @ 86f8ac01
......@@ -92,7 +92,6 @@ struct Epilogue
static_cast<int>(prob.C),
static_cast<int>(prob.Y),
static_cast<int>(prob.X)};
ck::Array<ck::index_t, 5> d_lengths = {};
ck::Array<ck::index_t, 5> in_strides{static_cast<int>(prob.C),
static_cast<int>(prob.Hi * prob.Wi * prob.G * prob.C),
......@@ -109,7 +108,6 @@ struct Epilogue
1,
static_cast<int>(prob.X * prob.C),
static_cast<int>(prob.C)};
ck::Array<ck::index_t, 5> d_strides = {};
ck::Array<ck::index_t, 2> conv_filter_strides = {2, 2};
ck::Array<ck::index_t, 2> conv_filter_dilations = {1, 1};
......
codegen/test/grouped_conv_fwd_multiple_d_v2.cpp
View file @ 86f8ac01
......@@ -92,7 +92,6 @@ struct Epilogue
static_cast<int>(prob.C),
static_cast<int>(prob.Y),
static_cast<int>(prob.X)};
ck::Array<ck::index_t, 5> d_lengths = {};
ck::Array<ck::index_t, 5> in_strides{static_cast<int>(prob.C),
static_cast<int>(prob.Hi * prob.Wi * prob.G * prob.C),
......@@ -109,7 +108,6 @@ struct Epilogue
1,
static_cast<int>(prob.X * prob.C),
static_cast<int>(prob.C)};
ck::Array<ck::index_t, 5> d_strides = {};
ck::Array<ck::index_t, 2> conv_filter_strides = {1, 1};
ck::Array<ck::index_t, 2> conv_filter_dilations = {1, 1};
......
codegen/test/grouped_conv_fwd_multiple_d_v3.cpp
View file @ 86f8ac01
......@@ -92,7 +92,6 @@ struct Epilogue
static_cast<int>(prob.C),
static_cast<int>(prob.Y),
static_cast<int>(prob.X)};
ck::Array<ck::index_t, 5> d_lengths = {};
ck::Array<ck::index_t, 5> in_strides{static_cast<int>(prob.C),
static_cast<int>(prob.Hi * prob.Wi * prob.G * prob.C),
......@@ -109,7 +108,6 @@ struct Epilogue
1,
static_cast<int>(prob.X * prob.C),
static_cast<int>(prob.C)};
ck::Array<ck::index_t, 5> d_strides = {};
ck::Array<ck::index_t, 2> conv_filter_strides = {2, 2};
ck::Array<ck::index_t, 2> conv_filter_dilations = {1, 1};
......
codegen/test/grouped_conv_fwd_multiple_d_v4.cpp
View file @ 86f8ac01
......@@ -92,7 +92,6 @@ struct Epilogue
static_cast<int>(prob.C),
static_cast<int>(prob.Y),
static_cast<int>(prob.X)};
ck::Array<ck::index_t, 5> d_lengths = {};
ck::Array<ck::index_t, 5> in_strides{static_cast<int>(prob.C),
static_cast<int>(prob.Hi * prob.Wi * prob.G * prob.C),
......@@ -109,7 +108,6 @@ struct Epilogue
1,
static_cast<int>(prob.X * prob.C),
static_cast<int>(prob.C)};
ck::Array<ck::index_t, 5> d_strides = {};
ck::Array<ck::index_t, 2> conv_filter_strides = {1, 1};
ck::Array<ck::index_t, 2> conv_filter_dilations = {1, 1};
......
codegen/test/rtc/src/kernel.cpp
View file @ 86f8ac01
......@@ -118,4 +118,4 @@ void kernel::launch(hipStream_t stream,
launch_kernel(impl->fun, stream, global, local, kernargs.data(), size);
}
} // namespace rtc
\ No newline at end of file
} // namespace rtc
codegen/test/rtc/src/tmp_dir.cpp
View file @ 86f8ac01
......@@ -45,4 +45,4 @@ void tmp_dir::execute(const std::string& cmd) const
tmp_dir::~tmp_dir() { std::filesystem::remove_all(this->path); }
} // namespace rtc
\ No newline at end of file
} // namespace rtc
docs/sphinx/requirements.in
View file @ 86f8ac01
rocm-docs-core==1.6.1
rocm-docs-core==1.6.2
sphinxcontrib-bibtex==2.6.2
docs/sphinx/requirements.txt
View file @ 86f8ac01
......@@ -103,7 +103,7 @@ requests==2.32.3
# via
# pygithub
# sphinx
rocm-docs-core==1.6.1
rocm-docs-core==1.6.2
# via -r requirements.in
six==1.16.0
# via pybtex
......
example/ck_tile/01_fmha/script/run_full_test.sh 0 → 100755
View file @ 86f8ac01
#!/bin/bash
#
# in order to run this script you'd first need to build the tile_example_fmha_fwd and tile_eaxmple_fmha_bwd executables in ../build/bin/
#
# run the script as "./run_full_test.sh <tag for your test environment> <branch name> <host name> <gpu_arch>
# input arguments:
# 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')
# host name : $hostname
# gpu architecture: e.g., gfx90a, or gfx942, etc.
#get the command line arguments:
export env_type=$1
echo 'Environment type: ' $env_type
export branch=$2
echo 'Branch name: ' $branch
export host_name=$3
echo 'Host name: ' $host_name
export GPU_arch=$4
echo 'GPU_arch: ' $GPU_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' >> $1;
}
#run verification tests
example/ck_tile/01_fmha/script/smoke_test_fwd.sh
example/ck_tile/01_fmha/script/smoke_test_bwd.sh
#run performance benchmarks
export fmha_fwd_log="perf_fmha_fwd_$GPU_arch.log"
print_log_header $fmha_fwd_log $env_type $branch $host_name
example/ck_tile/01_fmha/script/benchmark_fwd.sh 2>&1 | tee -a $fmha_fwd_log
export fmha_bwd_log="perf_fmha_bwd_$GPU_arch.log"
print_log_header $fmha_bwd_log $env_type $branch $host_name
example/ck_tile/01_fmha/script/benchmark_bwd.sh 2>&1 | tee -a $fmha_bwd_log
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View file @ 86f8ac01
......@@ -102,9 +102,10 @@ __global__ void
// offset base pointer for each work-group
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.z);
const long_index_t e_group_offset =
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const auto& ds_group_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
......@@ -117,14 +118,14 @@ __global__ void
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_group_offset[i]; });
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
if constexpr(isMultiA || isMultiB)
{
AsPointer p_as_grid_grp;
BsPointer p_bs_grid_grp;
const auto& as_group_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
const auto& as_batch_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
// compute_ptr_offset_of_n_ not need BatchStrideB so
// in case of MultiA is false but isMultiB is true
......@@ -135,27 +136,27 @@ __global__ void
static constexpr index_t NumATensor = AGridDesc_AK0_M_AK1::Size();
static_for<0, NumATensor, 1>{}([&](auto i) {
p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + as_n_offset[i];
p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + as_n_offset[i];
});
}
else
{
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
static_for<0, 1, 1>{}(
[&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + a_n_offset; });
[&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + a_n_offset; });
}
const auto& bs_group_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
const auto& bs_batch_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
static constexpr index_t NumBTensor = BGridDesc_BK0_N_BK1::Size();
static_for<0, NumBTensor, 1>{}(
[&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_group_offset[i]; });
[&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_batch_offset[i]; });
GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid_grp,
p_bs_grid_grp,
p_ds_grid_grp,
p_e_grid + e_group_offset + e_n_offset,
p_e_grid + e_batch_offset + e_n_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -168,19 +169,19 @@ __global__ void
}
else
{
const long_index_t a_group_offset =
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_group_offset =
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid + a_group_offset + a_n_offset,
p_bs_grid + b_group_offset,
p_as_grid + a_batch_offset + a_n_offset,
p_bs_grid + b_batch_offset,
p_ds_grid_grp,
p_e_grid + e_group_offset + e_n_offset,
p_e_grid + e_batch_offset + e_n_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -282,8 +283,7 @@ template <index_t NDimSpatial,
// in tuple for MultiAB), unpack if tuple was
// passed
typename BComputeDataType = AComputeDataType,
LoopScheduler LoopSched = make_default_loop_scheduler(),
index_t NumGroupsToMerge = 1>
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
: public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
ALayout,
......@@ -302,8 +302,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{
using DeviceOp = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle;
static_assert(NumGroupsToMerge >= 1);
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
......@@ -319,9 +317,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
using ConvToGemmFwdTransformer = TransformConvFwdToGemm<NDimSpatial,
ConvForwardSpecialization,
true /*SplitN*/,
ALayout,
ELayout,
NumGroupsToMerge>;
ADataType,
EDataType>;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
......@@ -520,8 +517,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{
static_for<0, NumATensor, 1>{}([&](auto i) {
// Init compute_ptr_offset_of_groups_ for multiple AB
compute_ptr_offset_of_groups_.BatchStrideA_(i) =
a_g_n_c_wis_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideA_(i) = a_g_n_c_wis_strides[0];
// Use GemmADataType/GemmBDataType to iterate over tuple (even if passed data
// type is not tuple)
......@@ -549,8 +545,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
});
static_for<0, NumBTensor, 1>{}([&](auto i) {
// Init compute_ptr_offset_of_groups_ for multiple AB
compute_ptr_offset_of_groups_.BatchStrideB_(i) =
b_g_k_c_xs_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideB_(i) = b_g_k_c_xs_strides[0];
using DataType = remove_cvref_t<tuple_element_t<i.value, GemmBDataType>>;
// It is possible that one of the AB is a pointer and one is a tuple.
......@@ -570,10 +565,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
}
else
{
compute_ptr_offset_of_groups_.BatchStrideA_ =
a_g_n_c_wis_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideB_ =
b_g_k_c_xs_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideA_ = a_g_n_c_wis_strides[0];
compute_ptr_offset_of_groups_.BatchStrideB_ = b_g_k_c_xs_strides[0];
compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_;
// p_as and p_bs are pointers
......@@ -590,8 +583,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
// D batch stride
compute_ptr_offset_of_groups_.BatchStrideDs_(i) =
ds_g_n_k_wos_strides[i][0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
compute_ptr_offset_of_n_.BatchStrideDs_(i) =
ds_g_n_k_wos_strides[i][1] * conv_N_per_block_;
......@@ -610,7 +602,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
ds_grid_desc_m_n_(i) =
DeviceOp::MakeEGridDescriptor_M_N<DLayout>(conv_to_gemm_transformer_d);
});
compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0];
compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_;
// populate desc for Ds/E
......@@ -734,7 +726,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_;
const index_t gdx = arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
const index_t gdy = arg.num_group_ / NumGroupsToMerge;
const index_t gdy = arg.num_group_;
const index_t gdz = num_workgroups_per_Conv_N;
const auto K =
......@@ -858,10 +850,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{
namespace ctc = tensor_layout::convolution;
const index_t G = arg.b_g_k_c_xs_lengths_[I0];
const index_t K = arg.b_g_k_c_xs_lengths_[I1];
const index_t C = arg.b_g_k_c_xs_lengths_[I2];
// check device
if(get_device_name() == "gfx908")
{
......@@ -910,42 +898,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
}
}
}
else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::Filter3x3)
{
if(C != 1)
{
return false;
}
for(index_t i = 0; i < NDimSpatial; ++i)
{
const index_t filter_spatial_dim = arg.b_g_k_c_xs_lengths_[i + I3];
if(filter_spatial_dim != I3)
{
return false;
}
}
if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
{
return false;
}
}
if constexpr(NumGroupsToMerge > 1)
{
if(!(C == 1))
{
return false;
}
if(G % NumGroupsToMerge != 0)
{
return false;
}
if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
{
return false;
}
}
// check vector access of A
// FIXME: layout
......@@ -955,16 +907,11 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
is_same_v<ALayout, ctc::NDHWGC>)
{
// Check access per C
const index_t C = arg.a_g_n_c_wis_lengths_[2];
if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
{
// If not possible, check access per G
if(!(ABlockTransferSrcVectorDim == 1 && C == 1 &&
is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>() &&
G % ABlockTransferSrcScalarPerVector == 0))
{
return false;
}
return false;
}
}
else
......@@ -981,6 +928,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<BLayout, ctc::KZYXGC>)
{
const index_t C = arg.b_g_k_c_xs_lengths_[2];
if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
{
return false;
......@@ -1004,6 +953,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
{
const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
{
valid = false;
......@@ -1048,6 +999,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
is_same_v<ELayout, ctc::NDHWGK>)
{
const index_t K = arg.e_g_n_k_wos_lengths_[2];
if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
{
return false;
......@@ -1345,8 +1298,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
<< BBlockTransferSrcScalarPerVector << ", "
<< CDEBlockTransferScalarPerVector_NPerBlock << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", "
<< NumGroupsToMerge
<< CShuffleNXdlPerWavePerShuffle
<< ">";
// clang-format on
......
include/ck/utility/f8_utils.hpp
View file @ 86f8ac01
......@@ -44,7 +44,7 @@ __host__ __device__ Y run_cast_to_f8(X x, uint32_t rng)
// convert to bitwise
using T_bitwise = typename NumericUtils<X>::bitwise_type;
T_bitwise x_bitwise = *(reinterpret_cast<T_bitwise*>(&x));
T_bitwise x_bitwise = bit_cast<T_bitwise>(x);
// unpack the input, depends on datatype
head = x_bitwise & NumericUtils<X>::head_mask;
......@@ -196,18 +196,17 @@ __host__ __device__ Y run_cast_from_f8(X x)
// prepare the codes
constexpr X nan_code = 0x80;
Y Inf, NegInf, NaN, Neg0;
using T_bitwise = typename NumericUtils<Y>::bitwise_type;
using T_bitwise = typename NumericUtils<Y>::bitwise_type;
constexpr T_bitwise Inf_bitwise = NumericUtils<Y>::Inf;
constexpr T_bitwise NegInf_bitwise = NumericUtils<Y>::NegInf;
constexpr T_bitwise NaN_bitwise = NumericUtils<Y>::NaN;
constexpr T_bitwise Neg0_bitwise = NumericUtils<Y>::Neg0;
Inf = *(reinterpret_cast<const Y*>(&Inf_bitwise));
NegInf = *(reinterpret_cast<const Y*>(&NegInf_bitwise));
NaN = *(reinterpret_cast<const Y*>(&NaN_bitwise));
Neg0 = *(reinterpret_cast<const Y*>(&Neg0_bitwise));
constexpr Y Inf = bit_cast<Y>(Inf_bitwise);
constexpr Y NegInf = bit_cast<Y>(NegInf_bitwise);
constexpr Y NaN = bit_cast<Y>(NaN_bitwise);
constexpr Y Neg0 = bit_cast<Y>(Neg0_bitwise);
// check if x is 0.0
if(x == 0)
......@@ -240,7 +239,7 @@ __host__ __device__ Y run_cast_from_f8(X x)
{
retval = x;
retval <<= 8;
return *(reinterpret_cast<const Y*>(&retval));
return bit_cast<Y>(retval);
}
// subnormal input
......@@ -264,7 +263,7 @@ __host__ __device__ Y run_cast_from_f8(X x)
}
retval = (sign << (out_exp + out_mant)) | (exponent << out_mant) | mantissa;
return *(reinterpret_cast<const Y*>(&retval));
return bit_cast<Y>(retval);
}
} // namespace
......
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