Skip to content

GitLab

"driver/device_direct_convolution_1.cuh" did not exist on "d2a488ddeca43e1b6bf6f1c3ceb4abf067c49962"
Commit 72c9f129 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents 241c261f ded0d83d
Show whitespace changes
Inline Side-by-side
Showing with 1142 additions and 73 deletions
CMakeLists.txt
View file @ 72c9f129
......@@ -26,7 +26,23 @@ set(version 1.1.0)
project(composable_kernel VERSION ${version} LANGUAGES CXX HIP)
include(CTest)
find_package(Python3 3.6 COMPONENTS Interpreter REQUIRED)
# Usage: for customized Python location cmake -DCK_USE_ALTERNATIVE_PYTHON="/opt/Python-3.8.13/bin/python3.8"
# CK Codegen requires dataclass which is added in Python 3.7
# Python version 3.8 is required for general good practice as it is default for Ubuntu 20.04
if(NOT CK_USE_ALTERNATIVE_PYTHON)
find_package(Python3 3.8 COMPONENTS Interpreter REQUIRED)
else()
message("Using alternative python version")
set(EXTRA_PYTHON_PATH)
# this is overly restrictive, we may need to be more flexible on the following
string(REPLACE "/bin/python3.8" "" EXTRA_PYTHON_PATH "${CK_USE_ALTERNATIVE_PYTHON}")
message("alternative python path is: ${EXTRA_PYTHON_PATH}")
find_package(Python3 3.6 COMPONENTS Interpreter REQUIRED)
add_definitions(-DPython3_EXECUTABLE="${CK_USE_ALTERNATIVE_PYTHON}")
set(Python3_EXECUTABLE "${CK_USE_ALTERNATIVE_PYTHON}")
set(PYTHON_EXECUTABLE "${CK_USE_ALTERNATIVE_PYTHON}")
set(ENV{LD_LIBRARY_PATH} "${EXTRA_PYTHON_PATH}/lib:$ENV{LD_LIBRARY_PATH}")
endif()
list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
......@@ -62,8 +78,14 @@ if (DTYPES)
endif()
message("DTYPES macro set to ${DTYPES}")
else()
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_ALL_DTYPES "ON")
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8)
set(CK_ENABLE_INT8 "ON")
set(CK_ENABLE_FP16 "ON")
set(CK_ENABLE_FP32 "ON")
set(CK_ENABLE_FP64 "ON")
set(CK_ENABLE_BF16 "ON")
set(CK_ENABLE_FP8 "ON")
set(CK_ENABLE_BF8 "ON")
endif()
#for f8/bf8_t type
......@@ -182,12 +204,18 @@ endif()
configure_file(include/ck/config.h.in ${CMAKE_CURRENT_BINARY_DIR}/include/ck/config.h)
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 500723302)
check_cxx_compiler_flag("-fno-offload-uniform-block" HAS_NO_OFFLOAD_UNIFORM_BLOCK)
if(HAS_NO_OFFLOAD_UNIFORM_BLOCK)
message("Adding the fno-offload-uniform-block compiler flag")
add_compile_options(-fno-offload-uniform-block)
endif()
endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
check_cxx_compiler_flag("-mllvm -enable-post-misched=0" HAS_ENABLE_POST_MISCHED)
if(HAS_ENABLE_POST_MISCHED)
message("Adding the enable-post-misched=0 compiler flag")
add_compile_options("SHELL: -mllvm -enable-post-misched=0")
endif()
endif()
set(check-coerce)
check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce)
......@@ -541,12 +569,7 @@ if(NOT DEFINED INSTANCES_ONLY)
PACKAGE_NAME examples
)
add_subdirectory(example)
if(GPU_TARGETS MATCHES "gfx9" AND NOT INSTANCES_ONLY)
add_subdirectory(codegen)
endif()
if(BUILD_TESTING)
add_subdirectory(test)
endif()
rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel
......@@ -563,6 +586,10 @@ if(NOT DEFINED INSTANCES_ONLY)
endif()
endif()
if(NOT DEFINED PROFILER_ONLY AND (GPU_TARGETS MATCHES "gfx9" OR DEFINED INSTANCES_ONLY))
add_subdirectory(codegen)
endif()
#Create an interface target for the include only files and call it "composablekernels"
include(CMakePackageConfigHelpers)
......
Jenkinsfile
View file @ 72c9f129
......@@ -262,10 +262,19 @@ def cmake_build(Map conf=[:]){
// reduce parallelism when compiling, clang uses too much memory
def nt = nthreads()
def cmd
def setup_cmd
def build_cmd
def execute_cmd = conf.get("execute_cmd", "")
if(!setup_args.contains("NO_CK_BUILD")){
def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j${nt} ${config_targets}")
if (setup_args.contains("gfx90a") && params.NINJA_BUILD_TRACE){
echo "running ninja build trace"
setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake -G Ninja ${setup_args} .. ")
build_cmd = conf.get("build_cmd", "${build_envs} ninja -j${nt} ${config_targets}")
}
else{
setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j${nt} ${config_targets}")
}
cmd = conf.get("cmd", """
${setup_cmd}
${build_cmd}
......@@ -281,7 +290,19 @@ def cmake_build(Map conf=[:]){
echo cmd
dir("build"){
//build CK
sh cmd
//run tests
if(!setup_args.contains("NO_CK_BUILD")){
if (setup_args.contains("gfx90a") && params.NINJA_BUILD_TRACE){
sh "/ninjatracing/ninjatracing .ninja_log > ck_build_trace.json"
archiveArtifacts "ck_build_trace.json"
sh "ninja test"
}
else{
sh "make check"
}
}
}
// Only archive from master or develop
......@@ -426,8 +447,9 @@ def runCKProfiler(Map conf=[:]){
archiveArtifacts "perf_resnet50_N4.log"
archiveArtifacts "perf_batched_gemm.log"
archiveArtifacts "perf_grouped_gemm.log"
archiveArtifacts "perf_conv_fwd.log"
archiveArtifacts "perf_conv_bwd_data.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"
......@@ -439,8 +461,9 @@ def runCKProfiler(Map conf=[:]){
stash name: "perf_resnet50_N4.log"
stash name: "perf_batched_gemm.log"
stash name: "perf_grouped_gemm.log"
stash name: "perf_conv_fwd.log"
stash name: "perf_conv_bwd_data.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"
......@@ -541,7 +564,7 @@ def Build_CK(Map conf=[:]){
cmake_build(conf)
dir("build"){
//run tests and examples
sh 'make -j check'
//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
......@@ -648,8 +671,9 @@ def process_results(Map conf=[:]){
unstash "perf_resnet50_N4.log"
unstash "perf_batched_gemm.log"
unstash "perf_grouped_gemm.log"
unstash "perf_conv_fwd.log"
unstash "perf_conv_bwd_data.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"
......@@ -681,8 +705,8 @@ 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_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 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-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
pipeline {
......@@ -746,6 +770,10 @@ pipeline {
name: "RUN_PERFORMANCE_TESTS",
defaultValue: true,
description: "Run the performance tests (default: ON)")
booleanParam(
name: "RUN_GROUPED_CONV_LARGE_CASES_TESTS",
defaultValue: false,
description: "Run the grouped conv large cases tests (default: OFF)")
booleanParam(
name: "RUN_CK_TILE_TESTS",
defaultValue: false,
......@@ -758,7 +786,10 @@ pipeline {
name: "BUILD_GFX12",
defaultValue: false,
description: "Build CK and run tests on gfx12 (default: OFF)")
booleanParam(
name: "NINJA_BUILD_TRACE",
defaultValue: false,
description: "Generate a ninja build trace (default: OFF)")
}
environment{
dbuser = "${dbuser}"
......@@ -792,6 +823,7 @@ pipeline {
}
agent{ label rocmnode("nogpu") }
environment{
setup_args = "NO_CK_BUILD"
execute_cmd = "find .. -not -path \'*.git*\' -iname \'*.h\' \
-o -not -path \'*.git*\' -iname \'*.hpp\' \
-o -not -path \'*.git*\' -iname \'*.cpp\' \
......@@ -808,7 +840,7 @@ pipeline {
--file-filter=*.cpp --force --enable=all --output-file=ck_cppcheck.log"
}
steps{
buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
buildHipClangJobAndReboot(setup_args:setup_args, setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
archiveArtifacts "build/ck_cppcheck.log"
cleanWs()
}
......@@ -820,6 +852,7 @@ pipeline {
}
agent{ label rocmnode("nogpu") }
environment{
setup_args = "NO_CK_BUILD"
execute_cmd = "find .. -not -path \'*.git*\' -iname \'*.h\' \
-o -not -path \'*.git*\' -iname \'*.hpp\' \
-o -not -path \'*.git*\' -iname \'*.cpp\' \
......@@ -831,7 +864,31 @@ pipeline {
| xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-12 -style=file {} | diff - {}\'"
}
steps{
buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
buildHipClangJobAndReboot(setup_args:setup_args, setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
cleanWs()
}
}
}
}
stage("Run Grouped Conv Large Case Tests")
{
parallel
{
stage("Run Grouped Conv Large Case Tests on gfx90a")
{
when {
beforeAgent true
expression { params.RUN_GROUPED_CONV_LARGE_CASES_TESTS.toBoolean() }
}
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "NO_CK_BUILD"
execute_args = """ ../script/cmake-ck-dev.sh ../ gfx90a && \
make -j64 test_grouped_convnd_fwd_large_cases_xdl && \
./bin/test_grouped_convnd_fwd_large_cases_xdl"""
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
cleanWs()
}
}
......@@ -936,10 +993,10 @@ pipeline {
}
agent{ label rocmnode("gfx90a") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1100;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " """
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx1100;gfx90a" \
-DGPU_TARGETS="gfx90a" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
......@@ -1043,7 +1100,7 @@ pipeline {
options { retry(1) }
agent{ label rocmnode("gfx90a")}
environment{
setup_args = """ -DGPU_TARGETS="gfx90a" -DBUILD_DEV=On """
setup_args = "NO_CK_BUILD"
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......
client_example/07_grouped_convnd_fwd/CMakeLists.txt
View file @ 72c9f129
......@@ -5,17 +5,17 @@ if(GPU_TARGETS MATCHES "gfx9")
add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_conv_operations)
if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_fwd_fp8 grouped_conv3d_fwd_fp8.cpp)
target_link_libraries(client_grouped_conv3d_fwd_fp8 PRIVATE composable_kernel::device_conv_operations)
endif()
if((DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_fwd_bf8 grouped_conv3d_fwd_bf8.cpp)
target_link_libraries(client_grouped_conv3d_fwd_bf8 PRIVATE composable_kernel::device_conv_operations)
endif()
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_fwd_fp8_bf8 grouped_conv3d_fwd_fp8_bf8.cpp)
target_link_libraries(client_grouped_conv3d_fwd_fp8_bf8 PRIVATE composable_kernel::device_conv_operations)
......
client_example/10_grouped_convnd_bwd_data/CMakeLists.txt
View file @ 72c9f129
......@@ -4,5 +4,7 @@ target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::
add_executable(client_grouped_conv3d_bwd_data grouped_conv3d_bwd_data.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data PRIVATE composable_kernel::device_conv_operations)
add_executable(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 PRIVATE composable_kernel::device_conv_operations)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 PRIVATE composable_kernel::device_conv_operations)
endif()
\ No newline at end of file
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 72c9f129
......@@ -2,10 +2,13 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
endif()
\ No newline at end of file
client_example/16_convnd_fwd/CMakeLists.txt
View file @ 72c9f129
......@@ -4,7 +4,7 @@ if((DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
endif()
if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_conv3d_fwd_fp16_comp_fp8 conv3d_fwd_fp16_comp_fp8.cpp)
target_link_libraries(client_conv3d_fwd_fp16_comp_fp8 PRIVATE composable_kernel::device_conv_operations)
endif()
......
client_example/20_splitk_gemm/CMakeLists.txt
View file @ 72c9f129
if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES))
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_gemm_operations)
endif()
client_example/24_grouped_conv_activation/CMakeLists.txt
View file @ 72c9f129
......@@ -47,6 +47,22 @@ target_link_libraries(client_conv3d_fwd_convscale_add_fp8 PRIVATE composable_ker
add_executable(client_conv3d_fwd_convscale_relu_fp8
grouped_convnd_fwd_convscale_relu/conv3d_fwd_convscale_relu_fp8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_relu_fp8 PRIVATE composable_kernel::device_conv_operations)
# Fwd convscale + ReLU + AMAX
add_executable(client_conv3d_fwd_convscale_relu_amax_fp8
grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_relu_amax_fp8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_relu_amax_fp8
PRIVATE composable_kernel::device_conv_operations
composable_kernel::device_other_operations
composable_kernel::device_reduction_operations
utility)
# Fwd convscale + AMAX
add_executable(client_conv3d_fwd_convscale_amax_fp8
grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_amax_fp8.cpp)
target_link_libraries(client_conv3d_fwd_convscale_amax_fp8
PRIVATE composable_kernel::device_conv_operations
composable_kernel::device_other_operations
composable_kernel::device_reduction_operations
utility)
# Fwd convscale
add_executable(client_conv3d_fwd_convscale_fp8
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8.cpp)
......
client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_reduce/common.hpp 0 → 100644
View file @ 72c9f129
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <string>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/utility/type.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_relu.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/permute_scale.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
#include "ck/library/utility/host_tensor.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ConvScaleRelu = ck::tensor_operation::element_wise::ScaleScaleRelu;
using ConvScale = ck::tensor_operation::element_wise::ScaleScalePass;
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_;
};
template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
const std::size_t& ds_size)
{
// 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product> +
// + ds_size * <output tensor size> =>
// => <output tensor size> * ( 2 * C * <filter spatial lengths product> + ds_size) =>
// => G * N * K * <output spatial lengths product> * (2 * C * <filter spatial lengths product> +
// ds_size)
ck::index_t G = weights_lengths[0];
ck::index_t N = output_lengths[1];
ck::index_t K = weights_lengths[1];
ck::index_t C = weights_lengths[2];
return G * N * K *
std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
std::end(output_lengths),
static_cast<std::size_t>(1),
std::multiplies<>()) *
(ds_size + static_cast<std::size_t>(2) * C *
std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
std::end(weights_lengths),
static_cast<std::size_t>(1),
std::multiplies<>()));
}
template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t GetTensorSize(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths)
{
return std::accumulate(std::begin(lengths),
std::end(lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>());
}
template <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
{
// sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
return sizeof(InDataType) * GetTensorSize<NumDimSpatial>(input_lengths);
}
template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
{
// sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
return sizeof(WeiDataType) * GetTensorSize<NumDimSpatial>(weights_lengths);
}
template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
std::size_t
GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
{
// sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
return sizeof(OutDataType) * GetTensorSize<NumDimSpatial>(output_lengths);
}
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename ConvElementOp,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim = 3,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool ConvolutionScale(SimpleDeviceMem& in,
SimpleDeviceMem& wei,
SimpleDeviceMem& out,
ConvElementOp elementwise_op,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_strides);
template <typename InDataType,
typename OutDataType,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim = 3>
bool TensorScaleConvert(SimpleDeviceMem& in,
SimpleDeviceMem& out,
float scale_out,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides);
template <typename InDataType,
typename OutDataType,
ck::ReduceTensorOp ReduceOpId,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim = 3>
bool TensorFullReduction(SimpleDeviceMem& tensor,
SimpleDeviceMem& out_amax,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides);
template <ck::index_t NumDimSpatial,
typename InDataType,
typename WeiDataType,
typename ConvOutDataType,
typename OutDataType,
typename ConvElementOp,
ck::ReduceTensorOp ReduceOp,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumNonSpatialDim = 3,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_fwd_convscale_reduce(
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
{
namespace ctc = ck::tensor_layout::convolution;
static_assert(NumDimSpatial == 3 && ck::is_same_v<InLayout, ctc::NDHWGC> &&
ck::is_same_v<WeiLayout, ctc::GKZYXC> &&
ck::is_same_v<OutLayout, ctc::NDHWGK>,
"Unsupported configuration");
const ck::index_t G = in_lengths[4];
const ck::index_t N = in_lengths[0];
const ck::index_t K = wei_lengths[1];
const ck::index_t C = in_lengths[5];
const ck::index_t Z = wei_lengths[2];
const ck::index_t Y = wei_lengths[3];
const ck::index_t X = wei_lengths[4];
const ck::index_t Di = in_lengths[1];
const ck::index_t Hi = in_lengths[2];
const ck::index_t Wi = in_lengths[3];
const ck::index_t Do = out_lengths[1];
const ck::index_t Ho = out_lengths[2];
const ck::index_t Wo = out_lengths[3];
const std::size_t in_mem_size = sizeof(InDataType) * N * Di * Hi * Wi * G * C;
const std::size_t wei_mem_size = sizeof(WeiDataType) * G * K * Z * Y * X * C;
const std::size_t conv_out_mem_size = sizeof(ConvOutDataType) * N * Do * Ho * Wo * G * K;
const std::size_t out_mem_size = sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
SimpleDeviceMem in(in_mem_size);
SimpleDeviceMem wei(wei_mem_size);
SimpleDeviceMem conv_out(conv_out_mem_size);
SimpleDeviceMem out(out_mem_size);
float scale_in = float(std::rand()) / float(RAND_MAX);
float scale_wei = float(std::rand()) / float(RAND_MAX);
float scale_out = float(std::rand()) / float(RAND_MAX);
// We have NDHWGC/GKZYXC/NDHWGK (x, weight, y) in memory space.
// However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
// Hence, we need to adjust the order of strides.
const std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
const std::array<ck::index_t, NumDimSpatial + 3> input_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
const std::array<ck::index_t, NumDimSpatial + 3> weights_lengths{G, K, C, Z, Y, X};
const std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
const std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
const std::array<ck::index_t, NumDimSpatial + 3> output_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
/*
* FP8 Convolution with Scaling
*/
std::cout << "\n\nConvolution with scale Benchmarking:" << std::endl;
auto elementwise_op = ConvElementOp{ck::tensor_operation::element_wise::Scale{scale_in},
ck::tensor_operation::element_wise::Scale{scale_wei},
{}};
auto conv_ok = ConvolutionScale<InDataType,
WeiDataType,
ConvOutDataType,
ConvElementOp,
InLayout,
WeiLayout,
OutLayout,
NumDimSpatial>(in,
wei,
conv_out,
elementwise_op,
input_lengths,
input_strides,
weights_lengths,
weights_strides,
output_lengths,
output_strides);
if(!conv_ok)
return false;
/*
* Scale with output weight and convert to FP8
*/
std::cout << "\n\nElement-wise scale + convert Benchmarking:" << std::endl;
auto elem_wise_ok = TensorScaleConvert<ConvOutDataType, OutDataType, NumDimSpatial>(
conv_out, out, scale_out, output_lengths, output_strides);
if(!elem_wise_ok)
return false;
/*
* Compute AMAX
*/
std::cout << "\n\nAMAX Benchmarking:" << std::endl;
SimpleDeviceMem amax_device(sizeof(ConvOutDataType));
auto reduction_ok =
TensorFullReduction<ConvOutDataType,
ConvOutDataType,
ck::ReduceTensorOp::AMAX,
NumDimSpatial>(conv_out, amax_device, output_lengths, output_strides);
if(!reduction_ok)
return false;
return true;
}
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename ConvElementOp,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim,
typename AComputeType,
typename BComputeType>
bool ConvolutionScale(SimpleDeviceMem& in,
SimpleDeviceMem& wei,
SimpleDeviceMem& out,
ConvElementOp elementwise_op,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_strides,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_strides)
{
const std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
const std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
const std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
const std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
const auto in_mem_size = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
const auto wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
const auto out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
std::size_t ds_size = 2; // 2 element-wise scale multipliers
if constexpr(ck::is_same_v<ConvElementOp, ConvScaleRelu>)
{
ds_size += 1; // +1 element-wise relu
}
std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
std::size_t num_bytes =
in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + out_mem_size;
using ConvDeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
ConvElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto conv_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
ConvDeviceOp>::GetInstances();
std::cout << "found " << conv_ptrs.size() << " instances" << std::endl;
std::string conv_best_op_name;
int conv_best_op_id = -1;
float conv_best_avg_time = std::numeric_limits<float>::max();
float conv_best_gb_per_sec = 0;
float conv_best_tflops = 0;
// profile device operation instances
std::cout << "Run all convolution instances and do timing" << std::endl;
for(int i = 0; i < conv_ptrs.size(); ++i)
{
auto& op_ptr = conv_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
std::array<const void*, 0>{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
out_lengths,
out_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
elementwise_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > conv_best_tflops)
{
conv_best_op_id = i;
conv_best_op_name = op_name;
conv_best_avg_time = avg_time;
conv_best_gb_per_sec = gb_per_sec;
conv_best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(conv_best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return false;
}
std::cout << "Best Perf: " << std::setw(10) << conv_best_avg_time << " ms, " << conv_best_tflops
<< " TFlops, " << conv_best_gb_per_sec << " GB/s, " << conv_best_op_name << std::endl;
// run the best instance
{
auto& op_ptr = conv_ptrs[conv_best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
std::array<const void*, 0>{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
out_lengths,
out_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
elementwise_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return true;
}
template <typename InDataType,
typename OutDataType,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim>
bool TensorScaleConvert(SimpleDeviceMem& in,
SimpleDeviceMem& out,
float scale_out,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides)
{
const auto tensor_size = GetTensorSize<NumDimSpatial>(lengths);
const std::size_t in_mem_size = sizeof(InDataType) * tensor_size;
const std::size_t out_mem_size = sizeof(OutDataType) * tensor_size;
std::size_t flop = 2 * tensor_size; // element-wise scale + convert
std::size_t bytes =
in_mem_size + sizeof(float) + out_mem_size; // read from in, scale, write to out
using DeviceScaleConvert =
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<InDataType>,
ck::Tuple<OutDataType>,
ck::tensor_operation::element_wise::Scale,
NumDimSpatial + NumNonSpatialDim>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceScaleConvert>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all DeviceScaleConvert instances and do timing" << std::endl;
auto scale_convert = ck::tensor_operation::element_wise::Scale{scale_out};
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
{strides},
{strides},
{in.GetDeviceBuffer()},
{out.GetDeviceBuffer()},
scale_convert);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance found." << std::endl;
return false;
}
else
{
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
{strides},
{strides},
{in.GetDeviceBuffer()},
{out.GetDeviceBuffer()},
scale_convert);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return true;
}
template <typename InDataType,
typename OutDataType,
ck::ReduceTensorOp ReduceOpId,
ck::index_t NumDimSpatial,
ck::index_t NumNonSpatialDim>
bool TensorFullReduction(SimpleDeviceMem& tensor,
SimpleDeviceMem& out_amax,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides)
{
const auto spatial_dim_size = std::accumulate(std::next(std::begin(lengths), NumNonSpatialDim),
std::end(lengths),
static_cast<std::size_t>(1),
std::multiplies<>());
const auto tensor_size = GetTensorSize<NumDimSpatial>(lengths);
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
// Get the reduction operation
using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation =
typename ck::reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename ck::reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
InElementwiseOperation in_elementwise_op;
AccElementwiseOperation acc_elementwise_op;
std::tie(in_elementwise_op, acc_elementwise_op) =
ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(tensor_size));
std::array<ck::index_t, 1> reduce_out_lengths{1};
std::array<ck::index_t, 1> reduce_out_strides{1};
SimpleDeviceMem partial_reduce_tensor(sizeof(OutDataType) * spatial_dim_size);
std::array<ck::index_t, NumDimSpatial> reduce_part_lengths;
std::copy(std::next(std::begin(lengths), NumNonSpatialDim),
std::end(lengths),
std::begin(reduce_part_lengths));
std::array<ck::index_t, NumDimSpatial> reduce_part_strides;
copy(HostTensorDescriptor(reduce_part_lengths).GetStrides(), reduce_part_strides);
{
std::cout << "\nReduction of nonspatial dimensions:" << std::endl;
using DeviceOp =
ck::tensor_operation::device::DeviceReduce<InDataType,
OutDataType,
OutDataType,
NumDimSpatial + NumNonSpatialDim,
NumNonSpatialDim,
ReduceOperation,
InElementwiseOperation,
PassThrough,
true, // PropagateNan
false>; // OutputIndex
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;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
std::array<int, NumNonSpatialDim> reduce_dims;
std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NumNonSpatialDim-1
ck::index_t num_in_elements = tensor_size;
ck::index_t num_out_elements = spatial_dim_size;
// profile device operation instances
std::cout << "Run partial reduction 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(lengths,
strides,
reduce_part_lengths,
reduce_part_strides,
reduce_dims,
1.0,
0.0,
tensor.GetDeviceBuffer(),
nullptr,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
in_elementwise_op,
PassThrough{});
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 num_bytes =
num_in_elements * sizeof(InDataType) + num_out_elements * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(ave_time < best_ave_time)
{
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance found." << std::endl;
return false;
}
else
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best instance
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
strides,
reduce_part_lengths,
reduce_part_strides,
reduce_dims,
1.0,
0.0,
tensor.GetDeviceBuffer(),
nullptr,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
in_elementwise_op,
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
{
std::cout << "\nReduction of spatial dimensions:" << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceReduce<OutDataType,
OutDataType,
OutDataType,
NumDimSpatial,
NumDimSpatial,
ReduceOperation,
PassThrough,
AccElementwiseOperation,
true, // PropagateNan
false>; // OutputIndex
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;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
std::array<int, NumDimSpatial> reduce_dims;
std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NumDimSpatial-1
ck::index_t num_in_elements = spatial_dim_size;
ck::index_t num_out_elements = 1;
// profile device operation instances
std::cout << "Run final reduction 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(reduce_part_lengths,
reduce_part_strides,
reduce_out_lengths,
reduce_out_strides,
reduce_dims,
1.0,
0.0,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
out_amax.GetDeviceBuffer(),
nullptr,
PassThrough{},
acc_elementwise_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 num_bytes =
num_in_elements * sizeof(OutDataType) + num_out_elements * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(ave_time < best_ave_time)
{
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance found." << std::endl;
return false;
}
else
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best instance
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(reduce_part_lengths,
reduce_part_strides,
reduce_out_lengths,
reduce_out_strides,
reduce_dims,
1.0,
0.0,
partial_reduce_tensor.GetDeviceBuffer(),
nullptr,
out_amax.GetDeviceBuffer(),
nullptr,
PassThrough{},
acc_elementwise_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
return true;
}
client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_amax_fp8.cpp 0 → 100644
View file @ 72c9f129
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::f8_t;
using WeiDataType = ck::f8_t;
using CShuffleDataType = float;
using ConvOutDataType = float; // data type of convolution result
using OutDataType = ck::f8_t; // data type of final result
using AComputeDataType = ck::f8_t;
using BComputeDataType = ck::f8_t;
using ConvElementOp = ConvScale;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 64;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
int main()
{
return run_grouped_conv_fwd_convscale_reduce<NumDimSpatial,
InDataType,
WeiDataType,
ConvOutDataType,
OutDataType,
ConvElementOp,
ReduceOpId,
InLayout,
WeiLayout,
OutLayout,
3,
AComputeDataType,
BComputeDataType>(
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_relu_amax_fp8.cpp 0 → 100644
View file @ 72c9f129
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::f8_t;
using WeiDataType = ck::f8_t;
using CShuffleDataType = float;
using ConvOutDataType = float; // data type of convolution result
using OutDataType = ck::f8_t; // data type of final result
using AComputeDataType = ck::f8_t;
using BComputeDataType = ck::f8_t;
using ConvElementOp = ConvScaleRelu;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 64;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
int main()
{
return run_grouped_conv_fwd_convscale_reduce<NumDimSpatial,
InDataType,
WeiDataType,
ConvOutDataType,
OutDataType,
ConvElementOp,
ReduceOpId,
InLayout,
WeiLayout,
OutLayout,
3,
AComputeDataType,
BComputeDataType>(
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/CMakeLists.txt
View file @ 72c9f129
......@@ -34,8 +34,17 @@ if (DTYPES)
endif()
message("DTYPES macro set to ${DTYPES}")
else()
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_ALL_DTYPES "ON")
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_INT8 "ON")
set(CK_ENABLE_FP16 "ON")
set(CK_ENABLE_FP32 "ON")
set(CK_ENABLE_FP64 "ON")
set(CK_ENABLE_BF16 "ON")
if (GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_ENABLE_FP8 -DCK_ENABLE_BF8)
set(CK_ENABLE_FP8 "ON")
set(CK_ENABLE_BF8 "ON")
endif()
endif()
if (GPU_TARGETS)
......
codegen/CMakeLists.txt
View file @ 72c9f129
......@@ -27,6 +27,8 @@ file(GLOB_RECURSE KERNEL_FILES CONFIGURE_DEPENDS
add_embed_library(ck_headers ${KERNEL_FILES} RELATIVE ${CK_ROOT}/include)
file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
##message(STATUS "SOURCE_FILES: ${SOURCES}")
# TODO: Use object library
add_library(ck_host STATIC ${SOURCES})
target_link_libraries(ck_host PRIVATE ck_headers)
......@@ -48,6 +50,4 @@ rocm_install(
)
rocm_install(DIRECTORY include/ck DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
if(BUILD_TESTING)
add_subdirectory(test)
endif()
codegen/test/CMakeLists.txt
View file @ 72c9f129
list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
add_subdirectory(rtc)
file(GLOB TEST_SRCS CONFIGURE_DEPENDS *.cpp)
foreach(TEST_SRC ${TEST_SRCS})
if(NOT INSTANCES_ONLY)
foreach(TEST_SRC ${TEST_SRCS})
set_source_files_properties(${TEST_SRC} PROPERTIES LANGUAGE HIP)
get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
add_executable(test_host_${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())
target_include_directories(test_host_${BASE_NAME} PUBLIC ${CK_ROOT}/include)
target_include_directories(test_host_${BASE_NAME} PUBLIC ${CK_ROOT}/library/include)
endforeach()
add_executable(codegen_test_${BASE_NAME} ${TEST_SRC})
if(CK_USE_ALTERNATIVE_PYTHON)
target_link_options(codegen_test_${BASE_NAME} PRIVATE -lstdc++fs)
endif()
add_dependencies(codegen codegen_test_${BASE_NAME})
add_dependencies(tests codegen_test_${BASE_NAME})
add_dependencies(check codegen_test_${BASE_NAME})
add_test(NAME codegen_test_${BASE_NAME} COMMAND codegen_test_${BASE_NAME})
message("adding test codegen_test_${BASE_NAME}")
target_link_libraries(codegen_test_${BASE_NAME} ck_rtc ck_host)
target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/codegen/test/include)
target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/include)
target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/library/include)
endforeach()
endif()
codegen/test/rtc/CMakeLists.txt
View file @ 72c9f129
find_package(hip)
file(GLOB RTC_SOURCES CONFIGURE_DEPENDS src/*.cpp)
add_library(ck_rtc ${RTC_SOURCES})
target_include_directories(ck_rtc PUBLIC include)
......
codegen/test/rtc/include/rtc/compile_kernel.hpp
View file @ 72c9f129
......@@ -2,14 +2,14 @@
#define GUARD_HOST_TEST_RTC_INCLUDE_RTC_COMPILE_KERNEL
#include <rtc/kernel.hpp>
#include <filesystem>
#include <ck/filesystem.hpp>
#include <string>
namespace rtc {
struct src_file
{
std::filesystem::path path;
CK::fs::path path;
std::string_view content;
};
......
codegen/test/rtc/include/rtc/tmp_dir.hpp
View file @ 72c9f129
......@@ -2,13 +2,13 @@
#define GUARD_HOST_TEST_RTC_INCLUDE_RTC_TMP_DIR
#include <string>
#include <filesystem>
#include <ck/filesystem.hpp>
namespace rtc {
struct tmp_dir
{
std::filesystem::path path;
CK::fs::path path;
tmp_dir(const std::string& prefix = "");
void execute(const std::string& cmd) const;
......
codegen/test/rtc/src/compile_kernel.cpp
View file @ 72c9f129
......@@ -70,9 +70,9 @@ kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options
for(const auto& src : srcs)
{
std::filesystem::path full_path = td.path / src.path;
std::filesystem::path parent_path = full_path.parent_path();
std::filesystem::create_directories(parent_path);
CK::fs::path full_path = td.path / src.path;
CK::fs::path parent_path = full_path.parent_path();
CK::fs::create_directories(parent_path);
write_string(full_path.string(), src.content);
if(src.path.extension().string() == ".cpp")
{
......@@ -86,7 +86,7 @@ kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options
td.execute(compiler() + options.flags);
auto out_path = td.path / out;
if(not std::filesystem::exists(out_path))
if(not CK::fs::exists(out_path))
throw std::runtime_error("Output file missing: " + out);
auto obj = read_buffer(out_path.string());
......
codegen/test/rtc/src/tmp_dir.cpp
View file @ 72c9f129
......@@ -31,10 +31,10 @@ std::string unique_string(const std::string& prefix)
}
tmp_dir::tmp_dir(const std::string& prefix)
: path(std::filesystem::temp_directory_path() /
: path(CK::fs::temp_directory_path() /
unique_string(prefix.empty() ? "ck-rtc" : "ck-rtc-" + prefix))
{
std::filesystem::create_directories(this->path);
CK::fs::create_directories(this->path);
}
void tmp_dir::execute(const std::string& cmd) const
......@@ -43,6 +43,6 @@ void tmp_dir::execute(const std::string& cmd) const
std::system(s.c_str());
}
tmp_dir::~tmp_dir() { std::filesystem::remove_all(this->path); }
tmp_dir::~tmp_dir() { CK::fs::remove_all(this->path); }
} // namespace rtc
docs/sphinx/requirements.in
View file @ 72c9f129
rocm-docs-core==1.6.2
rocm-docs-core==1.7.2
sphinxcontrib-bibtex==2.6.2
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