Skip to content

GitLab

Commit 036c5234 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin/develop' into aosewski/ggemm_multi_d2

parents 22995e9a 7843a8a7
Show whitespace changes
Inline Side-by-side
Showing with 1201 additions and 110 deletions
.github/CODEOWNERS
View file @ 036c5234
* @zjing14 @junliume @illsilin @carlushuang @aosewski @yigex
# Documentation files
docs/* @ROCm/rocm-documentation
*.md @ROCm/rocm-documentation
*.rst @ROCm/rocm-documentation
docs/* @ROCm/rocm-documentation @zjing14 @junliume @illsilin @carlushuang @aosewski @yigex
*.md @ROCm/rocm-documentation @zjing14 @junliume @illsilin @carlushuang @aosewski @yigex
*.rst @ROCm/rocm-documentation @zjing14 @junliume @illsilin @carlushuang @aosewski @yigex
.readthedocs.yaml @ROCm/rocm-documentation @zjing14 @junliume @illsilin @carlushuang @aosewski @yigex
# Header directory for Doxygen documentation
library/include/* @ROCm/rocm-documentation
library/include/* @ROCm/rocm-documentation @zjing14 @junliume @illsilin @carlushuang @aosewski @yigex
.readthedocs.yaml
View file @ 036c5234
......@@ -15,4 +15,4 @@ python:
build:
os: ubuntu-22.04
tools:
python: "3.8"
python: "3.10"
CMakeLists.txt
View file @ 036c5234
......@@ -26,7 +26,7 @@ set(version 1.1.0)
project(composable_kernel VERSION ${version} LANGUAGES CXX)
include(CTest)
find_package(Python3 3.8 COMPONENTS Interpreter REQUIRED)
find_package(Python3 3.6 COMPONENTS Interpreter REQUIRED)
list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
......@@ -202,7 +202,7 @@ endif()
option(USE_BITINT_EXTENSION_INT4 "Whether to enable clang's BitInt extension to provide int4 data type." OFF)
option(USE_OPT_NAVI3X "Whether to enable LDS cumode and Wavefront32 mode for NAVI3X silicons." OFF)
option(USE_OPT_GFX11 "Whether to enable LDS cumode and Wavefront32 mode for GFX11 silicons." OFF)
if(USE_BITINT_EXTENSION_INT4)
add_compile_definitions(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
......@@ -210,10 +210,10 @@ if(USE_BITINT_EXTENSION_INT4)
message("CK compiled with USE_BITINT_EXTENSION_INT4 set to ${USE_BITINT_EXTENSION_INT4}")
endif()
if(USE_OPT_NAVI3X)
if(USE_OPT_GFX11)
add_compile_options(-mcumode)
add_compile_options(-mno-wavefrontsize64)
message("CK compiled with USE_OPT_NAVI3X set to ${USE_OPT_NAVI3X}")
message("CK compiled with USE_OPT_GFX11 set to ${USE_OPT_GFX11}")
endif()
## Threads
......
Jenkinsfile
View file @ 036c5234
......@@ -515,30 +515,25 @@ def Build_CK(Map conf=[:]){
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 24, unit: 'HOURS')
{
//check whether running on Navi or MI300 node
def navi_node = 0
def mi300_node = 0
//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') ){
navi_node = 1
echo "This is a Navi node"
}
if ( runShell('grep -n "gfx942" rocminfo.log') ){
mi300_node = 1
echo "This is MI300 node"
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
do_perf_tests = 1
echo "Stash profiler and run performance tests"
}
cmake_build(conf)
dir("build"){
//run tests and examples
sh 'make -j check'
if (params.RUN_PERFORMANCE_TESTS && navi_node == 0 && mi300_node == 0 ){
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 Navi or MI300 nodes
//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 && mi300_node == 0 ){
// build deb packages for all MI100/200/300 targets and prepare to export
if (params.RUN_FULL_QA && do_perf_tests == 0 ){
// build deb packages for all gfx9 targets and prepare to export
sh 'make -j package'
archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb'
archiveArtifacts artifacts: 'composablekernel-tests_*.deb'
......@@ -546,7 +541,7 @@ def Build_CK(Map conf=[:]){
stash name: "ckprofiler_0.2.0_amd64.deb"
}
}
if (params.hipTensor_test && navi_node == 0 ){
if (params.hipTensor_test && do_perf_tests == 0 ){
//build and test hipTensor
sh """#!/bin/bash
rm -rf "${params.hipTensor_branch}".zip
......@@ -660,7 +655,8 @@ def process_results(Map conf=[:]){
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.1;COMPILER_VERSION=
0 21 * * * % ROCMVERSION=6.1;COMPILER_VERSION=;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;COMPILER_COMMIT=;USE_SCCACHE=false
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;COMPILER_COMMIT=;USE_SCCACHE=false''' : ""
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;COMPILER_COMMIT=;USE_SCCACHE=false
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
pipeline {
agent none
......@@ -727,6 +723,10 @@ pipeline {
name: "RUN_CODEGEN_TESTS",
defaultValue: true,
description: "Run the codegen tests (default: ON)")
booleanParam(
name: "BUILD_INSTANCES_ONLY",
defaultValue: false,
description: "Test building instances for various architectures simultaneously (default: OFF)")
}
environment{
dbuser = "${dbuser}"
......@@ -809,22 +809,22 @@ pipeline {
{
parallel
{
stage("Run Codegen Tests on MI100/MI200")
stage("Run Codegen Tests on gfx90a")
{
when {
beforeAgent true
expression { params.RUN_CODEGEN_TESTS.toBoolean() }
}
options { retry(2) }
agent{ label rocmnode("gfx908 || gfx90a")}
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="gfx908;gfx90a" \
-DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " .. && make -j check"""
-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)
......@@ -837,30 +837,30 @@ pipeline {
{
parallel
{
stage("Build CK and run Tests on MI100/MI200/MI300")
stage("Build CK for all gfx9 targets")
{
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("gfx908 || gfx90a") }
agent{ label rocmnode("gfx90a") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install \
-DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" \
-DCMAKE_EXE_LINKER_FLAGS=" -L ${env.WORKSPACE}/script -T hip_fatbin_insert " \
-DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " """
-DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " .. && make -j """
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
cleanWs()
}
}
stage("Build CK and run Tests on MI300")
stage("Build CK and run Tests on gfx942")
{
when {
beforeAgent true
......@@ -868,45 +868,65 @@ pipeline {
}
agent{ label rocmnode("gfx942") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx942" -DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " """
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx942" -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx942" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " .. && make -j """
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
cleanWs()
}
}
stage("Build CK and run Tests on MI100/MI200")
stage("Build CK and run Tests on gfx90a")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
}
agent{ label rocmnode("gfx908 || gfx90a") }
agent{ label rocmnode("gfx90a") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " """
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx908;gfx90a" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -Xclang -mllvm -Xclang -enable-post-misched=0 -O3 " .. && make -j """
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
cleanWs()
}
}
stage("Build CK and run Tests on Navi21")
stage("Build CK instances for different targets")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
expression { params.BUILD_INSTANCES_ONLY.toBoolean() && !params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("navi21") }
agent{ label rocmnode("gfx90a") }
environment{
execute_args = """ cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER="${build_compiler()}" \
-D CMAKE_BUILD_TYPE=Release \
-D GPU_TARGETS="gfx90a;gfx1030;gfx1101" \
-D INSTANCES_ONLY=ON \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j32 """
}
steps{
buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
cleanWs()
}
}
stage("Build CK and run Tests on gfx1030")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
}
agent{ label rocmnode("gfx1030") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1030" -DDL_KERNELS=ON -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
......@@ -920,13 +940,13 @@ pipeline {
cleanWs()
}
}
stage("Build CK and run Tests on Navi32")
stage("Build CK and run Tests on gfx1101")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
}
agent{ label rocmnode("navi32") }
agent{ label rocmnode("gfx1101") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1101" -DDL_KERNELS=ON -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
......@@ -947,27 +967,11 @@ pipeline {
{
parallel
{
stage("Run ckProfiler: gfx90*")
{
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() && params.RUN_PERFORMANCE_TESTS.toBoolean() }
}
options { retry(2) }
agent{ label rocmnode("gfx908 || gfx90a")}
environment{
setup_args = """ -DGPU_TARGETS="gfx908;gfx90a" -DBUILD_DEV=On """
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
cleanWs()
}
}
stage("Run ckProfiler: gfx90a")
{
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() && params.RUN_PERFORMANCE_TESTS.toBoolean() }
expression { params.RUN_PERFORMANCE_TESTS.toBoolean() }
}
options { retry(2) }
agent{ label rocmnode("gfx90a")}
......
client_example/11_grouped_conv_bwd_weight/common.hpp
View file @ 036c5234
// 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 <iomanip>
......@@ -160,6 +160,10 @@ bool run_grouped_conv_bwd_weight(
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace_dev(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
......
client_example/25_wrapper/wrapper_img2col.cpp
View file @ 036c5234
......@@ -181,4 +181,3 @@ int main(int argc, char* argv[])
{1, 1, 1} /*filter_dilations*/);
return 0;
}
// MI100 Perf: 0.255178 ms, 1698.9 GB/s,
client_example/30_gemm_multi_abd/CMakeLists.txt → client_example/30_gemm_bf16Aint8B/CMakeLists.txt
View file @ 036c5234
......@@ -10,4 +10,7 @@ if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf
add_executable(client_gemm_bf16_i8_bf16 gemm_xdl_bf16_i8.cpp)
target_link_libraries(client_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_gemm_multiply_bf16_i8_bf16 gemm_xdl_multiply_bf16_i8.cpp)
target_link_libraries(client_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
endif()
client_example/30_gemm_multi_abd/gemm_bias_fastgelu_xdl_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_bias_fastgelu_xdl_bf16_i8.cpp
View file @ 036c5234
......@@ -38,19 +38,19 @@ using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Col;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......
client_example/30_gemm_multi_abd/gemm_bias_xdl_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_bias_xdl_bf16_i8.cpp
View file @ 036c5234
......@@ -36,7 +36,7 @@ using D0DataType = BF16;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = BF16;
using A0Layout = Col;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
......@@ -45,12 +45,12 @@ using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = Add;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......
client_example/30_gemm_multi_abd/gemm_xdl_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_xdl_bf16_i8.cpp
View file @ 036c5234
......@@ -37,19 +37,19 @@ using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Col;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......@@ -74,12 +74,12 @@ struct SimpleDeviceMem
int main(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 64;
ck::index_t N = 1024;
ck::index_t K = 512;
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
if(argc == 1)
......
client_example/30_gemm_multi_abd/gemm_xdl_gelu_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_xdl_gelu_bf16_i8.cpp
View file @ 036c5234
......@@ -37,19 +37,19 @@ using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Col;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = FastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......
client_example/30_gemm_bf16Aint8B/gemm_xdl_multiply_bf16_i8.cpp 0 → 100644
View file @ 036c5234
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using DsDataType = ck::Tuple<B1DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<B1Layout>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Multiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
// clang-format on
int main(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 7)
{
M = std::stoi(argv[1]);
N = std::stoi(argv[2]);
K = std::stoi(argv[3]);
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideE = std::stoi(argv[6]);
}
else
{
printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
exit(0);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
f_matrix_space_size(M, K, StrideA, A0Layout{}));
SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
f_matrix_space_size(K, N, StrideB, B0Layout{}));
SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 1;
constexpr ck::index_t NumDTensor = 1;
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
BsLayout,
DsLayout,
Row,
AsDataType,
BsDataType,
DsDataType,
BF16,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB},
std::array<ck::index_t, NumDTensor>{0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return 0;
}
client_example/31_grouped_gemm_multi_abd/CMakeLists.txt → client_example/31_grouped_gemm_bf16Aint8B/CMakeLists.txt
View file @ 036c5234
......@@ -4,4 +4,13 @@ if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf
add_executable(client_grouped_gemm_fastgelu_bf16_i8_bf16 grouped_gemm_fastgelu_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_multiply_bf16_i8_bf16 grouped_gemm_multiply_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_bf16_i8_bf16 grouped_gemm_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
endif()
client_example/31_grouped_gemm_multi_abd/grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp → client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp
View file @ 036c5234
......@@ -38,19 +38,19 @@ using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Col;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......
client_example/31_grouped_gemm_multi_abd/grouped_gemm_fastgelu_xdl_bf16_i8.cpp → client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_fastgelu_xdl_bf16_i8.cpp
View file @ 036c5234
......@@ -15,6 +15,8 @@
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -36,7 +38,7 @@ using D0DataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Col;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
......@@ -45,12 +47,12 @@ using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = FastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp 0 → 100644
View file @ 036c5234
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using B0DataType = I8;
using B1DataType = BF16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using DsDataType = ck::Tuple<B1DataType, D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using B0Layout = Row;
using B1Layout = B0Layout;
using D0Layout = Row;
using DsLayout = ck::Tuple<B0Layout, D0Layout>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MultiplyAddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
d0_tensors_device, c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
d0_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumDTensor = 2;
using GroupedGemmKernelArgument =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(A0DataType) * problem_size.Ms[i] * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
d0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
gemm_descs.push_back({problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
problem_size.stride_Cs[i],
{0, 0}});
grouped_gemm_kernel_args_.push_back(
{a0_tensors_device[i]->GetDeviceBuffer(),
b0_tensors_device[i]->GetDeviceBuffer(),
{b1_tensors_device[i]->GetDeviceBuffer(), d0_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
{0, 0},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<A0Layout,
B0Layout,
DsLayout,
ELayout,
A0DataType,
B0DataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
std::vector<const void*> p_As = {};
std::vector<const void*> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(1 + rand() % 1024);
problem_size.Ns.push_back(6144);
problem_size.Ks.push_back(4096);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
<< problem_size.Ks[i] << std::endl;
}
return !run_grouped_gemm(problem_size, config);
}
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_xdl_bf16_i8.cpp 0 → 100644
View file @ 036c5234
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using B0DataType = I8;
using B1DataType = BF16;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<B1DataType>;
using EDataType = BF16;
using A0Layout = Row;
using B0Layout = Row;
using B1Layout = B0Layout;
using D0Layout = Row;
using DsLayout = ck::Tuple<B1Layout>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Multiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumDTensor = 1;
using GroupedGemmKernelArgument =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(A0DataType) * problem_size.Ms[i] * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
gemm_descs.push_back({problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
problem_size.stride_Cs[i],
{0}});
grouped_gemm_kernel_args_.push_back({a0_tensors_device[i]->GetDeviceBuffer(),
b0_tensors_device[i]->GetDeviceBuffer(),
{b1_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
{0},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<A0Layout,
B0Layout,
DsLayout,
ELayout,
A0DataType,
B0DataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
std::vector<const void*> p_As = {};
std::vector<const void*> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(1 + rand() % 1024);
problem_size.Ns.push_back(4096);
problem_size.Ks.push_back(4096);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
<< problem_size.Ks[i] << std::endl;
}
return !run_grouped_gemm(problem_size, config);
}
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_xdl_bf16_i8.cpp 0 → 100644
View file @ 036c5234
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 0;
using GroupedGemmKernelArgument = ck::tensor_operation::device::
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
gemm_descs.push_back(
{sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {}, 1});
grouped_gemm_kernel_args_.push_back(
{std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
b1_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumDTensor>{},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
std::array<ck::index_t, NumDTensor>{},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<AsLayout,
BsLayout,
DsLayout,
Row,
AsDataType,
BsDataType,
DsDataType,
BF16,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
std::vector<std::array<const void*, NumATensor>> p_As = {};
std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
op_ptr->SetElementwiseOps(
argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(1 + rand() % 1024);
problem_size.Ns.push_back(4096);
problem_size.Ks.push_back(4096);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
<< problem_size.Ks[i] << std::endl;
}
return !run_grouped_gemm(problem_size, config);
}
docs/sphinx/requirements.in
View file @ 036c5234
rocm-docs-core==0.38.1
rocm-docs-core==1.1.1
sphinxcontrib-bibtex==2.6.2
docs/sphinx/requirements.txt
View file @ 036c5234
#
# This file is autogenerated by pip-compile with Python 3.8
# This file is autogenerated by pip-compile with Python 3.10
# by the following command:
#
# pip-compile requirements.in
......@@ -48,12 +48,6 @@ idna==3.4
# via requests
imagesize==1.4.1
# via sphinx
importlib-metadata==6.8.0
# via
# sphinx
# sphinxcontrib-bibtex
importlib-resources==6.1.0
# via rocm-docs-core
jinja2==3.1.2
# via
# myst-parser
......@@ -99,8 +93,6 @@ pyjwt[crypto]==2.6.0
# via pygithub
pynacl==1.5.0
# via pygithub
pytz==2023.3.post1
# via babel
pyyaml==6.0
# via
# myst-parser
......@@ -111,7 +103,7 @@ requests==2.31.0
# via
# pygithub
# sphinx
rocm-docs-core==0.38.1
rocm-docs-core==1.1.1
# via -r requirements.in
six==1.16.0
# via
......@@ -165,7 +157,3 @@ urllib3==1.26.18
# via requests
wrapt==1.15.0
# via deprecated
zipp==3.17.0
# via
# importlib-metadata
# importlib-resources
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