CMake file changes for macOS universal2 support (#8953)

This commit is contained in:
Changming Sun 2021-09-04 13:30:33 -07:00 committed by GitHub
parent a9776d1c70
commit 60c98a86b7
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
8 changed files with 468 additions and 352 deletions

View file

@ -168,6 +168,14 @@ set(ONNX_CUSTOM_PROTOC_EXECUTABLE "" CACHE STRING "Specify custom protoc executa
# pre-build python path
option(onnxruntime_PREBUILT_PYTORCH_PATH "Path to pytorch installation dir")
if(APPLE)
if(NOT CMAKE_OSX_ARCHITECTURES)
message("Building ONNX Runtime for ${CMAKE_HOST_SYSTEM_PROCESSOR}")
endif()
elseif(NOT WIN32 AND NOT APPLE)
message("Building ONNX Runtime for ${CMAKE_SYSTEM_PROCESSOR}")
endif()
# Single output director for all binaries
set (RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin CACHE PATH "Single output directory for all binaries.")
@ -507,6 +515,30 @@ endmacro()
#Set global compile flags for all the source code(including third_party code like protobuf)
#This section must be before any add_subdirectory, otherwise build may fail because /MD,/MT mismatch
if (MSVC)
enable_language(ASM_MASM)
if(CMAKE_GENERATOR_PLATFORM)
# Multi-platform generator
set(onnxruntime_target_platform ${CMAKE_GENERATOR_PLATFORM})
else()
set(onnxruntime_target_platform ${CMAKE_SYSTEM_PROCESSOR})
endif()
if(onnxruntime_target_platform STREQUAL "ARM64")
set(onnxruntime_target_platform "ARM64")
elseif(onnxruntime_target_platform STREQUAL "ARM64EC")
set(onnxruntime_target_platform "ARM64EC")
elseif(onnxruntime_target_platform STREQUAL "ARM" OR CMAKE_GENERATOR MATCHES "ARM")
set(onnxruntime_target_platform "ARM")
elseif(onnxruntime_target_platform STREQUAL "x64" OR onnxruntime_target_platform STREQUAL "x86_64" OR onnxruntime_target_platform STREQUAL "AMD64" OR CMAKE_GENERATOR MATCHES "Win64")
set(onnxruntime_target_platform "x64")
elseif(onnxruntime_target_platform STREQUAL "Win32" OR onnxruntime_target_platform STREQUAL "x86" OR onnxruntime_target_platform STREQUAL "i386" OR onnxruntime_target_platform STREQUAL "i686")
set(onnxruntime_target_platform "x86")
if(NOT onnxruntime_BUILD_WEBASSEMBLY)
message("Enabling SAFESEH for x86 build")
set(CMAKE_ASM_MASM_FLAGS "${CMAKE_ASM_MASM_FLAGS} /safeseh")
endif()
endif()
#Always enable exception handling, even for Windows ARM
if(NOT onnxruntime_DISABLE_EXCEPTIONS)
string(APPEND CMAKE_CXX_FLAGS " /EHsc")
@ -555,6 +587,9 @@ if (MSVC)
SET(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} /guard:cf")
endif()
else()
if(NOT APPLE)
set(onnxruntime_target_platform ${CMAKE_SYSTEM_PROCESSOR})
endif()
if(onnxruntime_BUILD_FOR_NATIVE_MACHINE)
string(APPEND CMAKE_CXX_FLAGS " -march=native -mtune=native")
string(APPEND CMAKE_C_FLAGS " -march=native -mtune=native")
@ -736,6 +771,17 @@ else()
endif()
if(TARGET protoc)
set_target_properties(protoc PROPERTIES FOLDER "External/Protobuf")
get_target_property(PROTOC_OSX_ARCH protoc OSX_ARCHITECTURES)
if(PROTOC_OSX_ARCH)
if (${CMAKE_HOST_SYSTEM_PROCESSOR} IN_LIST PROTOC_OSX_ARCH)
message("protoc can run")
else()
list(APPEND PROTOC_OSX_ARCH ${CMAKE_HOST_SYSTEM_PROCESSOR})
set_target_properties(protoc PROPERTIES OSX_ARCHITECTURES "${CMAKE_HOST_SYSTEM_PROCESSOR}")
set_target_properties(libprotoc PROPERTIES OSX_ARCHITECTURES "${PROTOC_OSX_ARCH}")
set_target_properties(libprotobuf PROPERTIES OSX_ARCHITECTURES "${PROTOC_OSX_ARCH}")
endif()
endif()
endif()
if (onnxruntime_USE_FULL_PROTOBUF)
set(PROTOBUF_LIB libprotobuf)
@ -810,31 +856,39 @@ if(NOT TARGET re2::re2)
set(RE2_INCLUDE_DIR ${REPO_ROOT}/cmake/external/re2)
endif()
# Adding pytorch CPU info library
# TODO!! need a better way to find out the supported architectures
set(TARGET_ARCH "${CMAKE_SYSTEM_PROCESSOR}")
if (NOT DEFINED TARGET_ARCH OR "${TARGET_ARCH}" STREQUAL "")
set(TARGET_ARCH "${CMAKE_OSX_ARCHITECTURES}")
endif()
set(CPUINFO_SUPPORTED TRUE)
if (NOT DEFINED TARGET_ARCH OR "${TARGET_ARCH}" STREQUAL "")
message(WARNING
"Target processor architecture is not specified. cpuinfo not included")
set(CPUINFO_SUPPORTED FALSE)
elseif(NOT TARGET_ARCH MATCHES "^(i[3-6]86|AMD64|x86(_64)?|armv[5-8].*|aarch64|arm64)$")
message(WARNING
"Target processor architecture \"${CPUINFO_TARGET_PROCESSOR}\" is not supported in cpuinfo. "
"cpuinfo not included.")
set(CPUINFO_SUPPORTED FALSE)
elseif(MSVC AND (( CMAKE_SYSTEM_PROCESSOR MATCHES "^(ARM.*|arm.*)$" ) OR (CMAKE_GENERATOR_PLATFORM MATCHES "^(ARM.*|arm.*)$" ) ))
message(WARNING
"Cpuinfo not included for compilation problems with Windows ARM.")
set(CPUINFO_SUPPORTED FALSE)
elseif(WINDOWS_STORE)
message(WARNING
"Cpuinfo not included in Windows Store builds")
set(CPUINFO_SUPPORTED FALSE)
list(LENGTH CMAKE_OSX_ARCHITECTURES CMAKE_OSX_ARCHITECTURES_LEN)
if(APPLE)
if(CMAKE_OSX_ARCHITECTURES_LEN LESS_EQUAL 1)
set(CPUINFO_SUPPORTED TRUE)
elseif(onnxruntime_BUILD_APPLE_FRAMEWORK)
# We stitch multiple static libraries together when onnxruntime_BUILD_APPLE_FRAMEWORK is true,
# but that would not work for universal static libraries
message(FATAL_ERROR "universal binary is not supported for apple framework")
endif()
else()
if(onnxruntime_BUILD_WEBASSEMBLY)
set(CPUINFO_SUPPORTED FALSE)
else()
set(CPUINFO_SUPPORTED TRUE)
endif()
if(WIN32)
# Exclude Windows ARM build and Windows Store
if(${onnxruntime_target_platform} MATCHES "^(ARM.*|arm.*)$" )
message(WARNING "Cpuinfo not included for compilation problems with Windows ARM.")
set(CPUINFO_SUPPORTED FALSE)
elseif(WINDOWS_STORE)
message(WARNING "Cpuinfo not included in Windows Store builds")
set(CPUINFO_SUPPORTED FALSE)
endif()
elseif(NOT ${onnxruntime_target_platform} MATCHES "^(i[3-6]86|AMD64|x86(_64)?|armv[5-8].*|aarch64|arm64)$")
message(WARNING
"Target processor architecture \"${onnxruntime_target_platform}\" is not supported in cpuinfo. "
"cpuinfo not included.")
set(CPUINFO_SUPPORTED FALSE)
endif()
endif()
# TODO do we have to add target_include_directories to each project that uses this?
@ -1162,6 +1216,9 @@ function(onnxruntime_add_shared_library_module target_name)
endif()
onnxruntime_configure_target(${target_name})
if (onnxruntime_target_platform STREQUAL "x86" AND NOT onnxruntime_BUILD_WEBASSEMBLY)
target_link_options(${target_name} PRIVATE /SAFESEH)
endif()
endfunction()
function(onnxruntime_add_executable target_name)
@ -1170,6 +1227,9 @@ function(onnxruntime_add_executable target_name)
endif()
add_executable(${target_name} ${ARGN})
onnxruntime_configure_target(${target_name})
if (onnxruntime_target_platform STREQUAL "x86" AND NOT onnxruntime_BUILD_WEBASSEMBLY)
target_link_options(${target_name} PRIVATE /SAFESEH)
endif()
endfunction()
function(onnxruntime_add_include_to_target dst_target)
@ -1813,7 +1873,7 @@ if (WINDOWS_STORE)
target_link_options(onnxruntime PRIVATE /DYNAMICBASE /NXCOMPAT /APPCONTAINER)
target_link_options(winml_dll PRIVATE /DYNAMICBASE /NXCOMPAT /APPCONTAINER)
if (onnxruntime_target_platform STREQUAL "x86")
if (onnxruntime_target_platform STREQUAL "x86" AND NOT onnxruntime_BUILD_WEBASSEMBLY)
target_link_options(onnxruntime PRIVATE /SAFESEH)
target_link_options(winml_dll PRIVATE /SAFESEH)
endif()

View file

@ -174,8 +174,8 @@ set(onnxruntime_INTERNAL_LIBRARIES
${onnxruntime_tvm_libs}
onnxruntime_framework
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_mlas
onnxruntime_flatbuffers
)

View file

@ -65,24 +65,6 @@ else()
endif()
endif()
if(CMAKE_GENERATOR_PLATFORM)
# Multi-platform generator
set(onnxruntime_target_platform ${CMAKE_GENERATOR_PLATFORM})
else()
set(onnxruntime_target_platform ${CMAKE_SYSTEM_PROCESSOR})
endif()
if(onnxruntime_target_platform STREQUAL "ARM64")
set(onnxruntime_target_platform "ARM64")
elseif(onnxruntime_target_platform STREQUAL "ARM64EC")
set(onnxruntime_target_platform "ARM64EC")
elseif(onnxruntime_target_platform STREQUAL "ARM" OR CMAKE_GENERATOR MATCHES "ARM")
set(onnxruntime_target_platform "ARM")
elseif(onnxruntime_target_platform STREQUAL "x64" OR onnxruntime_target_platform STREQUAL "x86_64" OR onnxruntime_target_platform STREQUAL "AMD64" OR CMAKE_GENERATOR MATCHES "Win64")
set(onnxruntime_target_platform "x64")
elseif(onnxruntime_target_platform STREQUAL "Win32" OR onnxruntime_target_platform STREQUAL "x86" OR onnxruntime_target_platform STREQUAL "i386" OR onnxruntime_target_platform STREQUAL "i686")
set(onnxruntime_target_platform "x86")
endif()
if(onnxruntime_target_platform STREQUAL "ARM64EC")
if (MSVC)
link_directories("$ENV{VCINSTALLDIR}/Tools/MSVC/$ENV{VCToolsVersion}/lib/ARM64EC")
@ -185,18 +167,11 @@ if(MSVC)
elseif(onnxruntime_target_platform STREQUAL "x86")
set(X86 TRUE)
endif()
elseif(NOT onnxruntime_BUILD_WEBASSEMBLY)
if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64")
set(ARM64 TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64e")
set(ARM64 TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "arm")
set(ARM TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64")
set(X86_64 TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "i386")
set(X86 TRUE)
elseif(APPLE)
if(CMAKE_OSX_ARCHITECTURES_LEN LESS_EQUAL 1)
set(X64 TRUE)
endif()
elseif(NOT onnxruntime_BUILD_WEBASSEMBLY)
if (CMAKE_SYSTEM_NAME STREQUAL "Android")
if (CMAKE_ANDROID_ARCH_ABI STREQUAL "armeabi-v7a")
set(ARM TRUE)

View file

@ -79,7 +79,24 @@ endif()
# Set platform and arch for packaging
# Checks the names set by MLAS on non-Windows platforms first
if (CMAKE_SYSTEM_NAME STREQUAL "Android")
if(APPLE)
get_target_property(ONNXRUNTIME4J_OSX_ARCH onnxruntime4j_jni OSX_ARCHITECTURES)
list(LENGTH ONNXRUNTIME4J_OSX_ARCH ONNXRUNTIME4J_OSX_ARCH_LEN)
if(ONNXRUNTIME4J_OSX_ARCH)
if(ONNXRUNTIME4J_OSX_ARCH_LEN LESS_EQUAL 1)
list(GET ONNXRUNTIME4J_OSX_ARCH 0 JNI_ARCH)
message("Set Java ARCH TO macOS/iOS ${JNI_ARCH}")
else()
message(FATAL_ERROR "Java is currently not supported for macOS universal")
endif()
else()
set(JNI_ARCH ${CMAKE_HOST_SYSTEM_PROCESSOR})
message("Set Java ARCH TO macOS/iOS ${JNI_ARCH}")
endif()
if(JNI_ARCH STREQUAL "x86_64")
set(JNI_ARCH x64)
endif()
elseif (CMAKE_SYSTEM_NAME STREQUAL "Android")
set(JNI_ARCH ${ANDROID_ABI})
elseif (ARM64)
set(JNI_ARCH aarch64)

View file

@ -1,65 +1,62 @@
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
set(mlas_common_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/platform.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/threading.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/sgemm.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qdwconv.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/convolve.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/pooling.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/transpose.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/reorder.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/snchwc.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/activate.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/logistic.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/tanh.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/erf.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/compute.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/quantize.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_default.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qladd.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qlmul.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qpostprocessor.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qlgavgpool.cpp
set(MLAS_SRC_DIR ${ONNXRUNTIME_ROOT}/core/mlas/lib)
onnxruntime_add_static_library(onnxruntime_mlas
${MLAS_SRC_DIR}/platform.cpp
${MLAS_SRC_DIR}/threading.cpp
${MLAS_SRC_DIR}/sgemm.cpp
${MLAS_SRC_DIR}/qgemm.cpp
${MLAS_SRC_DIR}/qdwconv.cpp
${MLAS_SRC_DIR}/convolve.cpp
${MLAS_SRC_DIR}/pooling.cpp
${MLAS_SRC_DIR}/transpose.cpp
${MLAS_SRC_DIR}/reorder.cpp
${MLAS_SRC_DIR}/snchwc.cpp
${MLAS_SRC_DIR}/activate.cpp
${MLAS_SRC_DIR}/logistic.cpp
${MLAS_SRC_DIR}/tanh.cpp
${MLAS_SRC_DIR}/erf.cpp
${MLAS_SRC_DIR}/compute.cpp
${MLAS_SRC_DIR}/quantize.cpp
${MLAS_SRC_DIR}/qgemm_kernel_default.cpp
${MLAS_SRC_DIR}/qladd.cpp
${MLAS_SRC_DIR}/qlmul.cpp
${MLAS_SRC_DIR}/qpostprocessor.cpp
${MLAS_SRC_DIR}/qlgavgpool.cpp
)
if (onnxruntime_BUILD_WEBASSEMBLY)
if (onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
file(GLOB_RECURSE mlas_platform_srcs
"${ONNXRUNTIME_ROOT}/core/mlas/lib/wasm_simd/*.cpp"
)
else()
file(GLOB_RECURSE mlas_platform_srcs
"${ONNXRUNTIME_ROOT}/core/mlas/lib/wasm/*.cpp"
)
endif()
elseif(MSVC)
set(ONNXRUNTIME_MLAS_LIBS onnxruntime_mlas)
#TODO: set MASM flags properly
function(setup_mlas_source_for_windows)
#The onnxruntime_target_platform variable was added by Windows AI team in onnxruntime_common.cmake
#Don't use it for other platforms.
if((onnxruntime_target_platform STREQUAL "ARM64") OR (onnxruntime_target_platform STREQUAL "ARM64EC"))
set(PREPROCESS_ARMASM_FLAGS "")
set(ARMASM_FLAGS "")
if(onnxruntime_target_platform STREQUAL "ARM64")
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_neon.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_udot.cpp
target_sources(onnxruntime_mlas PRIVATE
${MLAS_SRC_DIR}/qgemm_kernel_neon.cpp
${MLAS_SRC_DIR}/qgemm_kernel_udot.cpp
)
set(mlas_platform_preprocess_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm64/QgemmU8X8KernelNeon.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm64/QgemmS8S8KernelNeon.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm64/QgemmU8X8KernelUdot.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm64/SgemmKernelNeon.asm
${MLAS_SRC_DIR}/arm64/QgemmU8X8KernelNeon.asm
${MLAS_SRC_DIR}/arm64/QgemmS8S8KernelNeon.asm
${MLAS_SRC_DIR}/arm64/QgemmU8X8KernelUdot.asm
${MLAS_SRC_DIR}/arm64/SgemmKernelNeon.asm
)
else()
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_neon.cpp
target_sources(onnxruntime_mlas PRIVATE
${MLAS_SRC_DIR}/qgemm_kernel_neon.cpp
)
set(mlas_platform_preprocess_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm64ec/QgemmU8X8KernelNeon.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm64ec/SgemmKernelNeon.asm
${MLAS_SRC_DIR}/arm64ec/QgemmU8X8KernelNeon.asm
${MLAS_SRC_DIR}/arm64ec/SgemmKernelNeon.asm
)
string(APPEND PREPROCESS_ARMASM_FLAGS " /arm64EC")
@ -88,288 +85,355 @@ elseif(MSVC)
DEPENDS ${asm_filename}
BYPRODUCTS ${preprocess_filename}
)
list(APPEND mlas_platform_srcs ${obj_filename})
target_sources(onnxruntime_mlas PRIVATE ${obj_filename})
endforeach()
elseif(onnxruntime_target_platform STREQUAL "ARM")
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm/sgemmc.cpp
target_sources(onnxruntime_mlas PRIVATE
${MLAS_SRC_DIR}/arm/sgemmc.cpp
)
elseif(onnxruntime_target_platform STREQUAL "x64")
enable_language(ASM_MASM)
file(GLOB_RECURSE mlas_platform_srcs_avx CONFIGURE_DEPENDS
"${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx/*.cpp"
"${MLAS_SRC_DIR}/intrinsics/avx/*.cpp"
)
set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "/arch:AVX")
file(GLOB_RECURSE mlas_platform_srcs_avx2 CONFIGURE_DEPENDS
"${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx2/*.cpp"
"${MLAS_SRC_DIR}/intrinsics/avx2/*.cpp"
)
set_source_files_properties(${mlas_platform_srcs_avx2} PROPERTIES COMPILE_FLAGS "/arch:AVX2")
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/dgemm.cpp
target_sources(onnxruntime_mlas PRIVATE
${MLAS_SRC_DIR}/dgemm.cpp
${mlas_platform_srcs_avx}
${mlas_platform_srcs_avx2}
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_avx2.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_sse.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_sse41.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx512/quantize_avx512f.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemmU8S8KernelAvx2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemmU8U8KernelAvx2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemmU8X8KernelAvx2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemmU8X8KernelAvx512Core.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemvU8S8KernelAvx2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemvU8S8KernelAvx512Core.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemvU8S8KernelAvx512Vnni.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/QgemvU8S8KernelAvxVnni.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/DgemmKernelSse2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/DgemmKernelAvx.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/DgemmKernelFma3.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/DgemmKernelAvx512F.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SgemmKernelSse2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SgemmKernelAvx.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SgemmKernelM1Avx.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SgemmKernelFma3.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SgemmKernelAvx512F.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SconvKernelSse2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SconvKernelAvx.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SconvKernelFma3.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SconvKernelAvx512F.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SpoolKernelSse2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SpoolKernelAvx.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SpoolKernelAvx512F.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/sgemma.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/cvtfp16a.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/SoftmaxKernelAvx.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/TransKernelFma3.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/TransKernelAvx512F.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/LogisticKernelFma3.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/TanhKernelFma3.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/amd64/ErfKernelFma3.asm
${MLAS_SRC_DIR}/qgemm_kernel_avx2.cpp
${MLAS_SRC_DIR}/qgemm_kernel_sse.cpp
${MLAS_SRC_DIR}/qgemm_kernel_sse41.cpp
${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
${MLAS_SRC_DIR}/amd64/QgemmU8S8KernelAvx2.asm
${MLAS_SRC_DIR}/amd64/QgemmU8U8KernelAvx2.asm
${MLAS_SRC_DIR}/amd64/QgemmU8X8KernelAvx2.asm
${MLAS_SRC_DIR}/amd64/QgemmU8X8KernelAvx512Core.asm
${MLAS_SRC_DIR}/amd64/QgemvU8S8KernelAvx2.asm
${MLAS_SRC_DIR}/amd64/QgemvU8S8KernelAvx512Core.asm
${MLAS_SRC_DIR}/amd64/QgemvU8S8KernelAvx512Vnni.asm
${MLAS_SRC_DIR}/amd64/QgemvU8S8KernelAvxVnni.asm
${MLAS_SRC_DIR}/amd64/DgemmKernelSse2.asm
${MLAS_SRC_DIR}/amd64/DgemmKernelAvx.asm
${MLAS_SRC_DIR}/amd64/DgemmKernelFma3.asm
${MLAS_SRC_DIR}/amd64/DgemmKernelAvx512F.asm
${MLAS_SRC_DIR}/amd64/SgemmKernelSse2.asm
${MLAS_SRC_DIR}/amd64/SgemmKernelAvx.asm
${MLAS_SRC_DIR}/amd64/SgemmKernelM1Avx.asm
${MLAS_SRC_DIR}/amd64/SgemmKernelFma3.asm
${MLAS_SRC_DIR}/amd64/SgemmKernelAvx512F.asm
${MLAS_SRC_DIR}/amd64/SconvKernelSse2.asm
${MLAS_SRC_DIR}/amd64/SconvKernelAvx.asm
${MLAS_SRC_DIR}/amd64/SconvKernelFma3.asm
${MLAS_SRC_DIR}/amd64/SconvKernelAvx512F.asm
${MLAS_SRC_DIR}/amd64/SpoolKernelSse2.asm
${MLAS_SRC_DIR}/amd64/SpoolKernelAvx.asm
${MLAS_SRC_DIR}/amd64/SpoolKernelAvx512F.asm
${MLAS_SRC_DIR}/amd64/sgemma.asm
${MLAS_SRC_DIR}/amd64/cvtfp16a.asm
${MLAS_SRC_DIR}/amd64/SoftmaxKernelAvx.asm
${MLAS_SRC_DIR}/amd64/TransKernelFma3.asm
${MLAS_SRC_DIR}/amd64/TransKernelAvx512F.asm
${MLAS_SRC_DIR}/amd64/LogisticKernelFma3.asm
${MLAS_SRC_DIR}/amd64/TanhKernelFma3.asm
${MLAS_SRC_DIR}/amd64/ErfKernelFma3.asm
)
else()
enable_language(ASM_MASM)
set(CMAKE_ASM_MASM_FLAGS "${CMAKE_ASM_MASM_FLAGS} /safeseh")
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_sse.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_sse41.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/i386/SgemmKernelSse2.asm
${ONNXRUNTIME_ROOT}/core/mlas/lib/i386/SgemmKernelAvx.asm
target_sources(onnxruntime_mlas PRIVATE
${MLAS_SRC_DIR}/qgemm_kernel_sse.cpp
${MLAS_SRC_DIR}/qgemm_kernel_sse41.cpp
${MLAS_SRC_DIR}/i386/SgemmKernelSse2.asm
${MLAS_SRC_DIR}/i386/SgemmKernelAvx.asm
)
endif()
endfunction()
if (onnxruntime_BUILD_WEBASSEMBLY)
if (onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
file(GLOB_RECURSE mlas_platform_srcs
"${MLAS_SRC_DIR}/wasm_simd/*.cpp"
)
else()
file(GLOB_RECURSE mlas_platform_srcs
"${MLAS_SRC_DIR}/wasm/*.cpp"
)
endif()
target_sources(onnxruntime_mlas PRIVATE ${mlas_platform_srcs})
elseif(MSVC)
setup_mlas_source_for_windows()
else()
if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64")
set(ARM64 TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64e")
set(ARM64 TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "arm")
set(ARM TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64")
set(X86_64 TRUE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "i386")
set(X86 TRUE)
endif()
if (CMAKE_SYSTEM_NAME STREQUAL "Android")
if (CMAKE_ANDROID_ARCH_ABI STREQUAL "armeabi-v7a")
set(ARM TRUE)
elseif (CMAKE_ANDROID_ARCH_ABI STREQUAL "arm64-v8a")
set(ARM64 TRUE)
elseif (CMAKE_ANDROID_ARCH_ABI STREQUAL "x86_64")
set(X86_64 TRUE)
elseif (CMAKE_ANDROID_ARCH_ABI STREQUAL "x86")
set(X86 TRUE)
endif()
elseif(CMAKE_SYSTEM_NAME STREQUAL "iOS" OR CMAKE_SYSTEM_NAME STREQUAL "iOSCross")
set(IOS TRUE)
else()
execute_process(
COMMAND ${CMAKE_C_COMPILER} -dumpmachine
OUTPUT_VARIABLE dumpmachine_output
ERROR_QUIET
)
if(dumpmachine_output MATCHES "^arm64.*")
set(ARM64 TRUE)
elseif(dumpmachine_output MATCHES "^arm.*")
set(ARM TRUE)
elseif(dumpmachine_output MATCHES "^aarch64.*")
set(ARM64 TRUE)
elseif(dumpmachine_output MATCHES "^(powerpc.*|ppc.*)")
set(POWER TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i.86|x86?)$")
set(X86 TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|amd64)$")
set(X86_64 TRUE)
endif()
endif()
if(ARM)
enable_language(ASM)
if(APPLE)
get_target_property(ONNXRUNTIME_MLAS_OSX_ARCH onnxruntime_mlas OSX_ARCHITECTURES)
set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} -mfpu=neon")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon")
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch32/QgemmU8X8KernelNeon.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/arm/sgemmc.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_neon.cpp
)
elseif(ARM64)
enable_language(ASM)
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch64/QgemmU8X8KernelNeon.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch64/QgemmS8S8KernelNeon.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch64/QgemmU8X8KernelUdot.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch64/SgemmKernelNeon.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch64/SgemvKernelNeon.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_neon.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_udot.cpp
)
elseif(POWER)
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/power/SgemmKernelPower.cpp
)
check_cxx_compiler_flag("-mcpu=power10" HAS_POWER10)
if(HAS_POWER10)
set(CMAKE_REQUIRED_FLAGS "-mcpu=power10")
check_cxx_source_compiles("
#include <altivec.h>
int main() {
__vector_quad acc0;
__builtin_mma_xxsetaccz (&acc0);
return 0;
}"
COMPILES_P10
)
if(COMPILES_P10)
check_cxx_source_compiles("
#include <sys/auxv.h>
int main() {
unsigned long hwcap2 = getauxval(AT_HWCAP2);
bool HasP10 = ((hwcap2 & PPC_FEATURE2_MMA) && (hwcap2 & PPC_FEATURE2_ARCH_3_1));
return 0;
}"
HAS_P10_RUNTIME
)
if (HAS_P10_RUNTIME)
set_source_files_properties(${mlas_common_srcs} PROPERTIES COMPILE_FLAGS "-DPOWER10")
if(NOT ONNXRUNTIME_MLAS_OSX_ARCH)
set(ONNXRUNTIME_MLAS_OSX_ARCH ${CMAKE_HOST_SYSTEM_PROCESSOR})
endif()
foreach(OSX_ARCH ${ONNXRUNTIME_MLAS_OSX_ARCH})
if (OSX_ARCH STREQUAL "arm64")
set(ARM64 TRUE)
elseif (OSX_ARCH STREQUAL "arm64e")
set(ARM64 TRUE)
elseif (OSX_ARCH STREQUAL "arm")
set(ARM TRUE)
elseif (OSX_ARCH STREQUAL "x86_64")
set(X86_64 TRUE)
elseif (OSX_ARCH STREQUAL "i386")
set(X86 TRUE)
endif()
endforeach()
elseif(ANDROID)
if (CMAKE_ANDROID_ARCH_ABI STREQUAL "armeabi-v7a")
set(ARM TRUE)
elseif (CMAKE_ANDROID_ARCH_ABI STREQUAL "arm64-v8a")
set(ARM64 TRUE)
elseif (CMAKE_ANDROID_ARCH_ABI STREQUAL "x86_64")
set(X86_64 TRUE)
elseif (CMAKE_ANDROID_ARCH_ABI STREQUAL "x86")
set(X86 TRUE)
endif()
else()
#Linux/FreeBSD/PowerPC/...
#The value of CMAKE_SYSTEM_PROCESSOR should be from `uname -m`
#Example values:
#arm64v8/ubuntu -> aarch64
#arm32v6/alpine -> armv7l
#arm32v7/centos -> armv7l
#ppc64le/debian -> ppc64le
#s390x/ubuntu -> s390x
#ppc64le/busybox -> ppc64le
#arm64v8/ubuntu -> aarch64
#Android: armv7-a aarch64 i686 x86_64
#chasun: I don't think anyone uses 'arm64'
if(CMAKE_SYSTEM_PROCESSOR MATCHES "^arm64.*")
set(ARM64 TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^arm.*")
set(ARM TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^aarch64.*")
set(ARM64 TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(powerpc.*|ppc.*)")
set(POWER TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i.86|x86?)$")
set(X86 TRUE)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|amd64)$")
set(X86_64 TRUE)
endif()
set(mlas_platform_srcs_power10
${ONNXRUNTIME_ROOT}/core/mlas/lib/power/SgemmKernelPOWER10.cpp
)
set_source_files_properties(${mlas_platform_srcs_power10} PROPERTIES COMPILE_FLAGS "-O2 -mcpu=power10")
set(mlas_platform_srcs
${mlas_platform_srcs}
${mlas_platform_srcs_power10}
)
endif()
endif()
elseif(X86)
enable_language(ASM)
set(mlas_platform_srcs_sse2
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_sse.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86/SgemmKernelSse2.S
)
set_source_files_properties(${mlas_platform_srcs_sse2} PROPERTIES COMPILE_FLAGS "-msse2")
if(APPLE)
get_target_property(ONNXRUNTIME_MLAS_MACOSX_ARCH onnxruntime_mlas OSX_ARCHITECTURES)
endif()
list(LENGTH ONNXRUNTIME_MLAS_MACOSX_ARCH ONNXRUNTIME_MLAS_MACOSX_ARCH_LENGH)
if(ONNXRUNTIME_MLAS_MACOSX_ARCH_LENGH GREATER 1)
set(ONNXRUNTIME_MLAS_MULTI_ARCH TRUE)
endif()
#If ONNXRUNTIME_MLAS_MULTI_ARCH is true, we need to go through every if branch below
#and split MLAS to multiple static libraries.
#Otherwise, it works like if(...) elseif(...) elseif(...) endif()
set(MLAS_SOURCE_IS_NOT_SET 1)
if(ARM)
enable_language(ASM)
set(mlas_platform_srcs_avx
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86/SgemmKernelAvx.S
)
set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} -mfpu=neon")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon")
set(mlas_platform_srcs
${mlas_platform_srcs_sse2}
${mlas_platform_srcs_avx}
)
elseif(X86_64)
enable_language(ASM)
set(mlas_platform_srcs
${MLAS_SRC_DIR}/aarch32/QgemmU8X8KernelNeon.S
${MLAS_SRC_DIR}/arm/sgemmc.cpp
${MLAS_SRC_DIR}/qgemm_kernel_neon.cpp
)
if(NOT ONNXRUNTIME_MLAS_MULTI_ARCH)
set(MLAS_SOURCE_IS_NOT_SET 0)
endif()
endif()
if(ARM64 AND MLAS_SOURCE_IS_NOT_SET )
enable_language(ASM)
set(mlas_platform_srcs
${MLAS_SRC_DIR}/aarch64/QgemmU8X8KernelNeon.S
${MLAS_SRC_DIR}/aarch64/QgemmS8S8KernelNeon.S
${MLAS_SRC_DIR}/aarch64/QgemmU8X8KernelUdot.S
${MLAS_SRC_DIR}/aarch64/SgemmKernelNeon.S
${MLAS_SRC_DIR}/aarch64/SgemvKernelNeon.S
${MLAS_SRC_DIR}/qgemm_kernel_neon.cpp
${MLAS_SRC_DIR}/qgemm_kernel_udot.cpp
)
if(ONNXRUNTIME_MLAS_MULTI_ARCH)
onnxruntime_add_static_library(onnxruntime_mlas_arm64 ${mlas_platform_srcs})
set_target_properties(onnxruntime_mlas_arm64 PROPERTIES OSX_ARCHITECTURES "arm64")
list(APPEND ONNXRUNTIME_MLAS_LIBS onnxruntime_mlas_arm64)
set(mlas_platform_srcs )
else()
set(MLAS_SOURCE_IS_NOT_SET 0)
endif()
endif()
if(POWER AND MLAS_SOURCE_IS_NOT_SET)
set(mlas_platform_srcs
${MLAS_SRC_DIR}/power/SgemmKernelPower.cpp
)
check_cxx_compiler_flag("-mcpu=power10" HAS_POWER10)
if(HAS_POWER10)
set(CMAKE_REQUIRED_FLAGS "-mcpu=power10")
check_cxx_source_compiles("
#include <altivec.h>
int main() {
__vector_quad acc0;
__builtin_mma_xxsetaccz (&acc0);
return 0;
}"
COMPILES_P10
)
if(COMPILES_P10)
check_cxx_source_compiles("
#include <sys/auxv.h>
int main() {
unsigned long hwcap2 = getauxval(AT_HWCAP2);
bool HasP10 = ((hwcap2 & PPC_FEATURE2_MMA) && (hwcap2 & PPC_FEATURE2_ARCH_3_1));
return 0;
}"
HAS_P10_RUNTIME
)
if (HAS_P10_RUNTIME)
set_source_files_properties(${mlas_common_srcs} PROPERTIES COMPILE_FLAGS "-DPOWER10")
endif()
set(mlas_platform_srcs_power10
${MLAS_SRC_DIR}/power/SgemmKernelPOWER10.cpp
)
set_source_files_properties(${mlas_platform_srcs_power10} PROPERTIES COMPILE_FLAGS "-O2 -mcpu=power10")
set(mlas_platform_srcs
${mlas_platform_srcs}
${mlas_platform_srcs_power10}
)
endif()
endif()
if(NOT ONNXRUNTIME_MLAS_MULTI_ARCH)
set(MLAS_SOURCE_IS_NOT_SET 0)
endif()
endif()
if(X86 AND MLAS_SOURCE_IS_NOT_SET)
enable_language(ASM)
# Forward the flags for the minimum target platform version from the C
# compiler to the assembler. This works around CMakeASMCompiler.cmake.in
# not including the logic to set this flag for the assembler.
set(CMAKE_ASM${ASM_DIALECT}_OSX_DEPLOYMENT_TARGET_FLAG "${CMAKE_C_OSX_DEPLOYMENT_TARGET_FLAG}")
set(mlas_platform_srcs_sse2
${MLAS_SRC_DIR}/qgemm_kernel_sse.cpp
${MLAS_SRC_DIR}/x86/SgemmKernelSse2.S
)
set_source_files_properties(${mlas_platform_srcs_sse2} PROPERTIES COMPILE_FLAGS "-msse2")
# The LLVM assembler does not support the .arch directive to enable instruction
# set extensions and also doesn't support AVX-512F instructions without
# turning on support via command-line option. Group the sources by the
# instruction set extension and explicitly set the compiler flag as appropriate.
set(mlas_platform_srcs_avx
${MLAS_SRC_DIR}/x86/SgemmKernelAvx.S
)
set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
set(mlas_platform_srcs_sse2
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_sse.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/DgemmKernelSse2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelSse2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmTransposePackB16x4Sse2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SconvKernelSse2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SpoolKernelSse2.S
)
set_source_files_properties(${mlas_platform_srcs_sse2} PROPERTIES COMPILE_FLAGS "-msse2")
set(mlas_platform_srcs
${mlas_platform_srcs_sse2}
${mlas_platform_srcs_avx}
)
if(NOT ONNXRUNTIME_MLAS_MULTI_ARCH)
set(MLAS_SOURCE_IS_NOT_SET 0)
endif()
endif()
if(X86_64 AND MLAS_SOURCE_IS_NOT_SET)
enable_language(ASM)
set(mlas_platform_srcs_avx
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/DgemmKernelAvx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelAvx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelM1Avx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelM1TransposeBAvx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmTransposePackB16x4Avx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SconvKernelAvx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SpoolKernelAvx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SoftmaxKernelAvx.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx/min_max_elements.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
# Forward the flags for the minimum target platform version from the C
# compiler to the assembler. This works around CMakeASMCompiler.cmake.in
# not including the logic to set this flag for the assembler.
set(CMAKE_ASM${ASM_DIALECT}_OSX_DEPLOYMENT_TARGET_FLAG "${CMAKE_C_OSX_DEPLOYMENT_TARGET_FLAG}")
set(mlas_platform_srcs_avx2
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemmU8S8KernelAvx2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemvU8S8KernelAvx2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemmU8U8KernelAvx2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemvU8S8KernelAvxVnni.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemmU8X8KernelAvx2.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/DgemmKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SconvKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/TransKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/LogisticKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/TanhKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/ErfKernelFma3.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx2/qladd_avx2.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx2/qdwconv_avx2.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx2} PROPERTIES COMPILE_FLAGS "-mavx2 -mfma")
# The LLVM assembler does not support the .arch directive to enable instruction
# set extensions and also doesn't support AVX-512F instructions without
# turning on support via command-line option. Group the sources by the
# instruction set extension and explicitly set the compiler flag as appropriate.
set(mlas_platform_srcs_avx512f
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/DgemmKernelAvx512F.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelAvx512F.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SconvKernelAvx512F.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SpoolKernelAvx512F.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/TransKernelAvx512F.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/intrinsics/avx512/quantize_avx512f.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx512f} PROPERTIES COMPILE_FLAGS "-mavx512f")
set(mlas_platform_srcs_sse2
${MLAS_SRC_DIR}/qgemm_kernel_sse.cpp
${MLAS_SRC_DIR}/x86_64/DgemmKernelSse2.S
${MLAS_SRC_DIR}/x86_64/SgemmKernelSse2.S
${MLAS_SRC_DIR}/x86_64/SgemmTransposePackB16x4Sse2.S
${MLAS_SRC_DIR}/x86_64/SconvKernelSse2.S
${MLAS_SRC_DIR}/x86_64/SpoolKernelSse2.S
)
set_source_files_properties(${mlas_platform_srcs_sse2} PROPERTIES COMPILE_FLAGS "-msse2")
set(mlas_platform_srcs_avx512core
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemvU8S8KernelAvx512Core.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemvU8S8KernelAvx512Vnni.S
${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/QgemmU8X8KernelAvx512Core.S
)
set_source_files_properties(${mlas_platform_srcs_avx512core} PROPERTIES COMPILE_FLAGS "-mavx512bw -mavx512dq -mavx512vl")
set(mlas_platform_srcs_avx
${MLAS_SRC_DIR}/x86_64/DgemmKernelAvx.S
${MLAS_SRC_DIR}/x86_64/SgemmKernelAvx.S
${MLAS_SRC_DIR}/x86_64/SgemmKernelM1Avx.S
${MLAS_SRC_DIR}/x86_64/SgemmKernelM1TransposeBAvx.S
${MLAS_SRC_DIR}/x86_64/SgemmTransposePackB16x4Avx.S
${MLAS_SRC_DIR}/x86_64/SconvKernelAvx.S
${MLAS_SRC_DIR}/x86_64/SpoolKernelAvx.S
${MLAS_SRC_DIR}/x86_64/SoftmaxKernelAvx.S
${MLAS_SRC_DIR}/intrinsics/avx/min_max_elements.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
set(mlas_platform_srcs
${ONNXRUNTIME_ROOT}/core/mlas/lib/dgemm.cpp
${ONNXRUNTIME_ROOT}/core/mlas/lib/qgemm_kernel_avx2.cpp
${mlas_platform_srcs_sse2}
${mlas_platform_srcs_avx}
${mlas_platform_srcs_avx2}
${mlas_platform_srcs_avx512f}
${mlas_platform_srcs_avx512core}
)
endif()
set(mlas_platform_srcs_avx2
${MLAS_SRC_DIR}/x86_64/QgemmU8S8KernelAvx2.S
${MLAS_SRC_DIR}/x86_64/QgemvU8S8KernelAvx2.S
${MLAS_SRC_DIR}/x86_64/QgemmU8U8KernelAvx2.S
${MLAS_SRC_DIR}/x86_64/QgemvU8S8KernelAvxVnni.S
${MLAS_SRC_DIR}/x86_64/QgemmU8X8KernelAvx2.S
${MLAS_SRC_DIR}/x86_64/DgemmKernelFma3.S
${MLAS_SRC_DIR}/x86_64/SgemmKernelFma3.S
${MLAS_SRC_DIR}/x86_64/SconvKernelFma3.S
${MLAS_SRC_DIR}/x86_64/TransKernelFma3.S
${MLAS_SRC_DIR}/x86_64/LogisticKernelFma3.S
${MLAS_SRC_DIR}/x86_64/TanhKernelFma3.S
${MLAS_SRC_DIR}/x86_64/ErfKernelFma3.S
${MLAS_SRC_DIR}/intrinsics/avx2/qladd_avx2.cpp
${MLAS_SRC_DIR}/intrinsics/avx2/qdwconv_avx2.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx2} PROPERTIES COMPILE_FLAGS "-mavx2 -mfma")
set(mlas_platform_srcs_avx512f
${MLAS_SRC_DIR}/x86_64/DgemmKernelAvx512F.S
${MLAS_SRC_DIR}/x86_64/SgemmKernelAvx512F.S
${MLAS_SRC_DIR}/x86_64/SconvKernelAvx512F.S
${MLAS_SRC_DIR}/x86_64/SpoolKernelAvx512F.S
${MLAS_SRC_DIR}/x86_64/TransKernelAvx512F.S
${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx512f} PROPERTIES COMPILE_FLAGS "-mavx512f")
set(mlas_platform_srcs_avx512core
${MLAS_SRC_DIR}/x86_64/QgemvU8S8KernelAvx512Core.S
${MLAS_SRC_DIR}/x86_64/QgemvU8S8KernelAvx512Vnni.S
${MLAS_SRC_DIR}/x86_64/QgemmU8X8KernelAvx512Core.S
)
set_source_files_properties(${mlas_platform_srcs_avx512core} PROPERTIES COMPILE_FLAGS "-mavx512bw -mavx512dq -mavx512vl")
set(mlas_platform_srcs
${MLAS_SRC_DIR}/dgemm.cpp
${MLAS_SRC_DIR}/qgemm_kernel_avx2.cpp
${mlas_platform_srcs_sse2}
${mlas_platform_srcs_avx}
${mlas_platform_srcs_avx2}
${mlas_platform_srcs_avx512f}
${mlas_platform_srcs_avx512core}
)
if(ONNXRUNTIME_MLAS_MULTI_ARCH)
onnxruntime_add_static_library(onnxruntime_mlas_x86_64 ${mlas_platform_srcs})
set_target_properties(onnxruntime_mlas_x86_64 PROPERTIES OSX_ARCHITECTURES "x86_64")
list(APPEND ONNXRUNTIME_MLAS_LIBS onnxruntime_mlas_x86_64)
set(mlas_platform_srcs )
else()
set(MLAS_SOURCE_IS_NOT_SET 1)
endif()
endif()
target_sources(onnxruntime_mlas PRIVATE ${mlas_platform_srcs})
endif()
onnxruntime_add_static_library(onnxruntime_mlas ${mlas_common_srcs} ${mlas_platform_srcs})
target_include_directories(onnxruntime_mlas PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc ${ONNXRUNTIME_ROOT}/core/mlas/lib)
foreach(mlas_target ${ONNXRUNTIME_MLAS_LIBS})
target_include_directories(${mlas_target} PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc ${MLAS_SRC_DIR})
endforeach()
set_target_properties(onnxruntime_mlas PROPERTIES FOLDER "ONNXRuntime")
if (WIN32)
target_compile_options(onnxruntime_mlas PRIVATE "/wd6385" "/wd4127")

View file

@ -134,8 +134,8 @@ target_link_libraries(onnxruntime_pybind11_state PRIVATE
onnxruntime_framework
onnxruntime_util
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_mlas
onnxruntime_flatbuffers
${pybind11_lib}
)

View file

@ -132,8 +132,8 @@ if (onnxruntime_BUILD_UNIT_TESTS)
list(APPEND ONNXRUNTIME_LIBS
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_mlas
onnxruntime_flatbuffers
)

View file

@ -410,8 +410,8 @@ set(onnxruntime_test_framework_libs
onnxruntime_framework
onnxruntime_util
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_mlas
)
set(onnxruntime_test_server_libs
@ -504,8 +504,8 @@ set(ONNXRUNTIME_TEST_LIBS
onnxruntime_framework
onnxruntime_util
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_mlas
onnxruntime_flatbuffers
)
@ -853,9 +853,9 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
file(GLOB_RECURSE MLAS_BENCH_SOURCE_FILES "${MLAS_BENCH_DIR}/*.cpp" "${MLAS_BENCH_DIR}/*.h")
onnxruntime_add_executable(onnxruntime_mlas_benchmark ${MLAS_BENCH_SOURCE_FILES})
target_include_directories(onnxruntime_mlas_benchmark PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc)
target_link_libraries(onnxruntime_mlas_benchmark PRIVATE benchmark::benchmark onnxruntime_util onnxruntime_framework onnxruntime_mlas onnxruntime_common ${CMAKE_DL_LIBS})
target_link_libraries(onnxruntime_mlas_benchmark PRIVATE benchmark::benchmark onnxruntime_util onnxruntime_framework ${ONNXRUNTIME_MLAS_LIBS} onnxruntime_common ${CMAKE_DL_LIBS})
if(NOT WIN32)
target_link_libraries(onnxruntime_mlas_benchmark PRIVATE nsync_cpp)
target_link_libraries(onnxruntime_mlas_benchmark PRIVATE nsync_cpp ${CMAKE_DL_LIBS})
endif()
set_target_properties(onnxruntime_mlas_benchmark PROPERTIES FOLDER "ONNXRuntimeTest")
endif()
@ -896,8 +896,8 @@ if(onnxruntime_ENABLE_EAGER_MODE)
onnxruntime_framework
flatbuffers
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_mlas
onnx
onnx_proto
${PROTOBUF_LIB}
@ -1105,7 +1105,7 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
endif()
target_include_directories(onnxruntime_mlas_test PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc ${ONNXRUNTIME_ROOT}
${CMAKE_CURRENT_BINARY_DIR})
set(onnxruntime_mlas_test_libs GTest::gtest GTest::gmock onnxruntime_mlas onnxruntime_common)
set(onnxruntime_mlas_test_libs GTest::gtest GTest::gmock ${ONNXRUNTIME_MLAS_LIBS} onnxruntime_common)
if(NOT WIN32)
list(APPEND onnxruntime_mlas_test_libs nsync_cpp ${CMAKE_DL_LIBS})
endif()