Skip to content

GitLab

Commit 54d3e2f1 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into ck_tile/moe

parents 199f7f71 b8addae2
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1855 additions and 49 deletions
CMakeLists.txt
View file @ 54d3e2f1
......@@ -553,12 +553,7 @@ 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
......@@ -575,6 +570,10 @@ 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 @ 54d3e2f1
......@@ -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
......@@ -543,7 +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
......@@ -684,8 +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 {
......@@ -765,7 +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}"
......@@ -799,6 +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\' \
......@@ -815,7 +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()
}
......@@ -827,6 +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\' \
......@@ -838,7 +864,7 @@ 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()
}
}
......@@ -967,10 +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 """
}
......@@ -1074,7 +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/24_grouped_conv_activation/CMakeLists.txt
View file @ 54d3e2f1
......@@ -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/conv3d_fwd_convscale_amax_fp8.cpp 0 → 100644
View file @ 54d3e2f1
// 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 @ 54d3e2f1
// 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;
}
codegen/CMakeLists.txt
View file @ 54d3e2f1
......@@ -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 @ 54d3e2f1
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} ${TEST_SRC})
add_dependencies(codegen codegen_test_${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 @ 54d3e2f1
find_package(hip)
file(GLOB RTC_SOURCES CONFIGURE_DEPENDS src/*.cpp)
add_library(ck_rtc ${RTC_SOURCES})
target_include_directories(ck_rtc PUBLIC include)
......
docs/sphinx/requirements.in
View file @ 54d3e2f1
rocm-docs-core==1.7.0
rocm-docs-core==1.7.2
sphinxcontrib-bibtex==2.6.2
docs/sphinx/requirements.txt
View file @ 54d3e2f1
......@@ -103,7 +103,7 @@ requests==2.32.3
# via
# pygithub
# sphinx
rocm-docs-core==1.7.0
rocm-docs-core==1.7.2
# via -r requirements.in
six==1.16.0
# via pybtex
......
example/01_gemm/gemm_xdl_fp8.cpp
View file @ 54d3e2f1
// 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 "common.hpp"
......@@ -7,7 +7,7 @@
using ADataType = ck::f8_t;
using BDataType = ck::f8_t;
using CDataType = ck::half_t;
using CDataType = ck::f8_t;
using AccDataType = float;
using CShuffleDataType = float;
......
example/01_gemm/run_gemm_example.inc
View file @ 54d3e2f1
// 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.
#pragma once
......@@ -34,11 +34,11 @@ inline __host__ __device__ constexpr double get_rtol()
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 1e-1; // 240 and 224 are acceptable
return 2e-1;
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 1.5e-1; // 57344 and 49152 are acceptable
return 2e-1;
}
else
{
......@@ -75,11 +75,11 @@ inline __host__ __device__ constexpr double get_atol()
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 16.1; // 240 and 224 are acceptable
return 2e-1;
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 8192.1; // 57344 and 49152 are acceptable
return 2e-1;
}
else
{
......
example/62_convnd_activ/CMakeLists.txt
View file @ 54d3e2f1
......@@ -3,6 +3,7 @@ add_subdirectory(convinvscale)
add_subdirectory(convscale)
add_subdirectory(convscale_relu)
add_subdirectory(convscale_add)
add_subdirectory(convscale_reduce)
add_subdirectory(multi_AB)
add_subdirectory(unary)
......
example/62_convnd_activ/convscale_reduce/CMakeLists.txt 0 → 100644
View file @ 54d3e2f1
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_convnd_activ_xdl_convscale_reduce)
add_example_executable(example_convnd_fwd_xdl_convscale_relu_amax_fp8 convnd_fwd_xdl_convscale_relu_amax_fp8.cpp)
add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_relu_amax_fp8)
add_example_executable(example_convnd_fwd_xdl_convscale_amax_fp8 convnd_fwd_xdl_convscale_amax_fp8.cpp)
add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_amax_fp8)
set(target 1)
endif()
endforeach()
example/62_convnd_activ/convscale_reduce/convnd_fwd_xdl_convscale_amax_fp8.cpp 0 → 100644
View file @ 54d3e2f1
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_convscale_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
using InDataType = ck::f8_t;
using WeiDataType = ck::f8_t;
using AccDataType = float;
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;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ConvScale;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<>,
ConvOutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 32, 1, 8>,
8,
AComputeDataType,
BComputeDataType>;
#include "run_convnd_fwd_example.inc"
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/62_convnd_activ/convscale_reduce/convnd_fwd_xdl_convscale_relu_amax_fp8.cpp 0 → 100644
View file @ 54d3e2f1
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_convscale_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
using InDataType = ck::f8_t;
using WeiDataType = ck::f8_t;
using AccDataType = float;
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;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ConvScaleRelu;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<>,
ConvOutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 32, 1, 8>,
8,
AComputeDataType,
BComputeDataType>;
#include "run_convnd_fwd_example.inc"
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/62_convnd_activ/convscale_reduce/run_convnd_fwd_example.inc 0 → 100644
View file @ 54d3e2f1
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
bool run_convnd_fwd_example(int argc, char* argv[])
{
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
// instantiate in and wei element ops, will
// instantiate out_element_op below for every iteration
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) {
constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
using InLayout = decltype(in_layout);
using WeiLayout = decltype(wei_layout);
using OutLayout = decltype(out_layout);
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
ndim_spatial_value,
InDataType,
WeiDataType,
ConvOutDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op);
};
namespace ctc = ck::tensor_layout::convolution;
if(conv_param.num_dim_spatial_ == 1)
{
return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{});
}
else if(conv_param.num_dim_spatial_ == 2)
{
return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{});
}
else if(conv_param.num_dim_spatial_ == 3)
{
return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
}
return true;
}
example/ck_tile/01_fmha/CMakeLists.txt
View file @ 54d3e2f1
# generate a list of kernels, but not actually emit files at config stage
# validate user-specified fmha_fwd API list
set(FMHA_FWD_KNOWN_APIS "fwd;fwd_splitkv;fwd_appendkv")
set(FMHA_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${FMHA_FWD_KNOWN_APIS}) & link, or \"all\".")
if(FMHA_FWD_ENABLE_APIS STREQUAL "all")
set(FMHA_FWD_ENABLE_APIS ${FMHA_FWD_KNOWN_APIS})
endif()
foreach(api ${FMHA_FWD_ENABLE_APIS})
if(NOT "${api}" IN_LIST FMHA_FWD_KNOWN_APIS)
message(FATAL_ERROR "${api} isn't a known api: ${FMHA_FWD_KNOWN_APIS}.")
endif()
endforeach()
# "fwd" is a must-have api for the fmha_fwd example, add it if not specified
if(NOT "fwd" IN_LIST FMHA_FWD_ENABLE_APIS)
list(APPEND FMHA_FWD_ENABLE_APIS "fwd")
endif()
string(REPLACE ";" "," FMHA_FWD_APIS "${FMHA_FWD_ENABLE_APIS}")
# generate a list of kernels, but not actually emit files at config sta
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api fwd,fwd_splitkv --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/fwd_blob_list.txt
--api ${FMHA_FWD_APIS} --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/fwd_blob_list.txt
)
execute_process(
......@@ -17,7 +37,7 @@ file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt FMHA_BWD_GEN_BLOBS)
add_custom_command(
OUTPUT ${FMHA_FWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api fwd,fwd_splitkv --output_dir ${CMAKE_CURRENT_BINARY_DIR}
--api ${FMHA_FWD_APIS} --output_dir ${CMAKE_CURRENT_BINARY_DIR}
)
add_custom_command(
......@@ -60,6 +80,20 @@ else()
endif()
list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-undefined-func-template -fgpu-flush-denormals-to-zero)
# conditionally enable call to the fwd_splitkv API in fmha_fwd example
if("fwd_splitkv" IN_LIST FMHA_FWD_ENABLE_APIS)
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_SPLITKV_API=1)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_SPLITKV_API=0)
endif()
# conditionally enable call to the fwd_appendkv API in fmha_fwd example
if("fwd_appendkv" IN_LIST FMHA_FWD_ENABLE_APIS)
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_APPENDKV_API=1)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_APPENDKV_API=0)
endif()
# Allow comparing floating points directly in order to check sentinel values
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-float-equal)
......
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