|
|
|
@ -1,266 +1,241 @@
|
|
|
|
|
# Clang or AppleClang (see CMP0025)
|
|
|
|
|
if(NOT DEFINED C_CLANG AND CMAKE_CXX_COMPILER_ID MATCHES "Clang")
|
|
|
|
|
include_guard(GLOBAL)
|
|
|
|
|
|
|
|
|
|
include(CheckCXXCompilerFlag)
|
|
|
|
|
include(CheckSymbolExists)
|
|
|
|
|
include(CMakePushCheckState)
|
|
|
|
|
|
|
|
|
|
set(TEST_DIR ${PROJECT_SOURCE_DIR}/cmake/test)
|
|
|
|
|
|
|
|
|
|
if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
|
|
|
|
|
set(C_CLANG 1)
|
|
|
|
|
elseif(NOT DEFINED C_GCC AND CMAKE_CXX_COMPILER_ID MATCHES "GNU")
|
|
|
|
|
elseif(MAKE_CXX_COMPILER_ID MATCHES "GNU")
|
|
|
|
|
set(C_GCC 1)
|
|
|
|
|
elseif(NOT DEFINED C_MSVC AND CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
|
|
|
|
|
elseif(CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
|
|
|
|
|
set(C_MSVC 1)
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
# Detect current compilation architecture and create standard definitions
|
|
|
|
|
include(CheckSymbolExists)
|
|
|
|
|
function(detect_architecture symbol arch)
|
|
|
|
|
macro(detect_architecture symbol arch)
|
|
|
|
|
if (NOT DEFINED ARCHITECTURE)
|
|
|
|
|
set(CMAKE_REQUIRED_QUIET 1)
|
|
|
|
|
check_symbol_exists("${symbol}" "" ARCHITECTURE_${arch})
|
|
|
|
|
unset(CMAKE_REQUIRED_QUIET)
|
|
|
|
|
|
|
|
|
|
# The output variable needs to be unique across invocations otherwise
|
|
|
|
|
# CMake's crazy scope rules will keep it defined
|
|
|
|
|
if (ARCHITECTURE_${arch})
|
|
|
|
|
set(ARCHITECTURE "${arch}" PARENT_SCOPE)
|
|
|
|
|
set(ARCHITECTURE_${arch} 1 PARENT_SCOPE)
|
|
|
|
|
add_definitions(-DARCHITECTURE_${arch}=1)
|
|
|
|
|
set(ARCHITECTURE ${arch})
|
|
|
|
|
set(ARCHITECTURE_${arch} TRUE)
|
|
|
|
|
add_compile_definitions(ARCHITECTURE_${arch})
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
endmacro()
|
|
|
|
|
|
|
|
|
|
macro(force_ext_available extension)
|
|
|
|
|
message(STATUS "Looking for __${extension}__ - forced found")
|
|
|
|
|
set(HAS_${extension} 1 CACHE INTERNAL "")
|
|
|
|
|
endmacro()
|
|
|
|
|
|
|
|
|
|
function(detect_extensions extension)
|
|
|
|
|
unset(HAS_${extension})
|
|
|
|
|
if (ARGC EQUAL 2 AND (${ARGV1})) # force available
|
|
|
|
|
force_ext_available(${extension})
|
|
|
|
|
endif()
|
|
|
|
|
check_symbol_exists("__${extension}__" "" HAS_${extension})
|
|
|
|
|
if (HAS_${extension})
|
|
|
|
|
add_compile_definitions(USE_${extension})
|
|
|
|
|
endif()
|
|
|
|
|
endfunction()
|
|
|
|
|
|
|
|
|
|
if (NOT ENABLE_GENERIC)
|
|
|
|
|
if (C_MSVC)
|
|
|
|
|
function(detect_msvc_native_opt)
|
|
|
|
|
try_run(RUN_AVX512 COMPILE_AVX512 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx512.cxx" COMPILE_DEFINITIONS /arch:AVX512)
|
|
|
|
|
if (COMPILE_AVX512 AND RUN_AVX512 EQUAL 0)
|
|
|
|
|
set(ARCH_OPT "AVX512" PARENT_SCOPE)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
try_run(RUN_AVX2 COMPILE_AVX2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx2.cxx" COMPILE_DEFINITIONS /arch:AVX2)
|
|
|
|
|
if (COMPILE_AVX2 AND RUN_AVX2 EQUAL 0)
|
|
|
|
|
set(ARCH_OPT "AVX2" PARENT_SCOPE)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
try_run(RUN_AVX COMPILE_AVX "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx.cxx" COMPILE_DEFINITIONS /arch:AVX)
|
|
|
|
|
if (COMPILE_AVX AND RUN_AVX EQUAL 0)
|
|
|
|
|
set(ARCH_OPT "AVX" PARENT_SCOPE)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
# Supporting 32-bit x86, what year is it?
|
|
|
|
|
set(COMPILE_DEF "")
|
|
|
|
|
set(ARCH_OPT "" PARENT_SCOPE)
|
|
|
|
|
if (ARCHITECTURE_x86)
|
|
|
|
|
set(COMPILE_DEF "/arch:SSE2")
|
|
|
|
|
set(ARCH_OPT "SSE2" PARENT_SCOPE)
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
try_run(RUN_SSE42 COMPILE_SSE42 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse42.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
|
|
|
|
|
if (COMPILE_SSE42 AND RUN_SSE42 EQUAL 0)
|
|
|
|
|
force_ext_available(SSE4_2)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
try_run(RUN_SSE41 COMPILE_SSE41 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse41.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
|
|
|
|
|
if (COMPILE_SSE41 AND RUN_SSE41 EQUAL 0)
|
|
|
|
|
force_ext_available(SSE4_1)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
try_run(RUN_SSSE3 COMPILE_SSSE3 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_ssse3.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
|
|
|
|
|
if (COMPILE_SSSE3 AND RUN_SSSE3 EQUAL 0)
|
|
|
|
|
force_ext_available(SSSE3)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
try_run(RUN_SSE3 COMPILE_SSE3 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse3.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
|
|
|
|
|
if (COMPILE_SSE3 AND RUN_SSE3 EQUAL 0)
|
|
|
|
|
force_ext_available(SSE3)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
try_run(RUN_SSE2 COMPILE_SSE2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse2.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
|
|
|
|
|
if (COMPILE_SSE2 AND RUN_SSE2 EQUAL 0)
|
|
|
|
|
force_ext_available(SSE2)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
if (ARCHITECTURE_x86)
|
|
|
|
|
# At this point we might as well...
|
|
|
|
|
set(ARCH_OPT "IA32" PARENT_SCOPE)
|
|
|
|
|
return()
|
|
|
|
|
endif()
|
|
|
|
|
endfunction()
|
|
|
|
|
|
|
|
|
|
if (C_MSVC)
|
|
|
|
|
detect_architecture("_M_AMD64" x86_64)
|
|
|
|
|
detect_architecture("_M_IX86" x86)
|
|
|
|
|
detect_architecture("_M_ARM" ARM)
|
|
|
|
|
detect_architecture("_M_ARM64" ARM64)
|
|
|
|
|
else()
|
|
|
|
|
else()
|
|
|
|
|
detect_architecture("__x86_64__" x86_64)
|
|
|
|
|
detect_architecture("__i386__" x86)
|
|
|
|
|
detect_architecture("__arm__" ARM)
|
|
|
|
|
detect_architecture("__aarch64__" ARM64)
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
if (NOT DEFINED ARCHITECTURE)
|
|
|
|
|
set(ARCHITECTURE "GENERIC")
|
|
|
|
|
set(ARCHITECTURE_GENERIC 1)
|
|
|
|
|
add_definitions(-DARCHITECTURE_GENERIC=1)
|
|
|
|
|
message(FATAL_ERROR "Not supported. Please add needed architecture detection.")
|
|
|
|
|
endif()
|
|
|
|
|
message(STATUS "Target architecture: ${ARCHITECTURE}")
|
|
|
|
|
|
|
|
|
|
set(TEST_DIR ${PROJECT_SOURCE_DIR}/cmake/test)
|
|
|
|
|
# Note: On x86 MSVC's /arch:SSE2 enables all SSE intrinsics support and is default option.
|
|
|
|
|
# On x86_64 MSVC's SSE is supported and enabled, so only AVX selection is needed.
|
|
|
|
|
|
|
|
|
|
# flag that set the minimum cpu flag requirements
|
|
|
|
|
# used to create re-distribuitable binary
|
|
|
|
|
if (${ARCHITECTURE} MATCHES "x86_64|x86" AND (FORCE_SSSE3 OR FORCE_SSE41))
|
|
|
|
|
if (FORCE_SSSE3)
|
|
|
|
|
set(HAS_SSSE3 ON CACHE BOOL "SSSE3 SIMD enabled")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mssse3" )
|
|
|
|
|
message(STATUS "Use SSSE3 SIMD instructions")
|
|
|
|
|
add_definitions(-DUSE_SSSE3)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
message(STATUS "Use MSVC SSSE3 SIMD instructions")
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_SSSE3)
|
|
|
|
|
if (FORCE_SSSE3)
|
|
|
|
|
message(WARNING "FORCE_SSSE3 flag is deprecated, please use ARCH_OPT option.")
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
if (C_MSVC)
|
|
|
|
|
if (ARCHITECTURE_x86)
|
|
|
|
|
set(ARCH_OPT "SSE2")
|
|
|
|
|
endif()
|
|
|
|
|
elseif (FORCE_SSE41)
|
|
|
|
|
set(HAS_SSSE3 ON CACHE BOOL "SSSE3 SIMD enabled")
|
|
|
|
|
set(HAS_SSE4_1 ON CACHE BOOL "Architecture has SSE 4.1 SIMD enabled")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -msse4.1" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -msse4.1" )
|
|
|
|
|
message(STATUS "Use SSE 4.1 SIMD instructions")
|
|
|
|
|
add_definitions(-DUSE_SSSE3)
|
|
|
|
|
add_definitions(-DUSE_SSE4_1)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
# seems that from MSVC 2015 comiler doesn't support those flags
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
message(STATUS "Use SSE 4.1 SIMD instructions")
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_SSSE3)
|
|
|
|
|
add_definitions(-DUSE_SSE4_1)
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
else ()
|
|
|
|
|
if (${ARCHITECTURE} MATCHES "x86_64|x86")
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_SSE2 COMPILE_SSE2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse2.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
force_ext_available(SSSE3)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_SSE2 COMPILE_SSE2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse2.cxx" COMPILE_DEFINITIONS -msse2 -O0)
|
|
|
|
|
set(FORCE_OPT "-mssse3")
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_SSE2 AND RUN_SSE2 EQUAL 0)
|
|
|
|
|
set(HAS_SSE2 ON CACHE BOOL "Architecture has SSSE2 SIMD enabled")
|
|
|
|
|
message(STATUS "Use SSE2 SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse2" )
|
|
|
|
|
add_definitions(-DUSE_SSE2)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_SSE2)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_SSE2 OFF CACHE BOOL "Architecture does not have SSSE2 SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_SSSE3 COMPILE_SSSE3 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_ssse3.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_SSSE3 COMPILE_SSSE3 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_ssse3.cxx" COMPILE_DEFINITIONS -mssse3 -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_SSSE3 AND RUN_SSSE3 EQUAL 0)
|
|
|
|
|
set(HAS_SSSE3 ON CACHE BOOL "Architecture has SSSE3 SIMD enabled")
|
|
|
|
|
message(STATUS "Use SSSE3 SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mssse3" )
|
|
|
|
|
add_definitions(-DUSE_SSSE3)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
# seems not present on MSVC 2017
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_SSSE3)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_SSSE3 OFF CACHE BOOL "Architecture does not have SSSE3 SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_SSE4_1 COMPILE_SSE4_1 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse41.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_SSE4_1 COMPILE_SSE4_1 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse41.cxx" COMPILE_DEFINITIONS -msse4.1 -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_SSE4_1 AND RUN_SSE4_1 EQUAL 0)
|
|
|
|
|
set(HAS_SSE4_1 ON CACHE BOOL "Architecture has SSE 4.1 SIMD enabled")
|
|
|
|
|
message(STATUS "Use SSE 4.1 SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -msse4.1" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -msse4.1" )
|
|
|
|
|
add_definitions(-DUSE_SSE4_1)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
# seems not present on MSVC 2017
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_SSE4_1)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_SSE4_1 OFF CACHE BOOL "Architecture does not have SSE 4.1 SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_SSE4_2 COMPILE_SSE4_2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse42.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_SSE4_2 COMPILE_SSE4_2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse42.cxx" COMPILE_DEFINITIONS -msse4.2 -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_SSE4_2 AND RUN_SSE4_2 EQUAL 0)
|
|
|
|
|
set(HAS_SSE4_2 ON CACHE BOOL "Architecture has SSE 4.2 SIMD enabled")
|
|
|
|
|
message(STATUS "Use SSE 4.2 SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -msse4.2" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -msse4.2" )
|
|
|
|
|
add_definitions(-DUSE_SSE4_2)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
# seems not present on MSVC 2017
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_SSE4_2)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_SSE4_2 OFF CACHE BOOL "Architecture does not have SSE 4.2 SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_AVX COMPILE_AVX "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_AVX COMPILE_AVX "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx.cxx" COMPILE_DEFINITIONS -mavx -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_AVX AND RUN_AVX EQUAL 0)
|
|
|
|
|
set(HAS_AVX ON CACHE BOOL "Architecture has AVX SIMD enabled")
|
|
|
|
|
message(STATUS "Use AVX SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mavx" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mavx" )
|
|
|
|
|
add_definitions(-DUSE_AVX)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /arch:AVX" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox /arch:AVX" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_AVX)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_AVX OFF CACHE BOOL "Architecture does not have AVX SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_AVX2 COMPILE_AVX2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx2.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_AVX2 COMPILE_AVX2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx2.cxx" COMPILE_DEFINITIONS -mavx2 -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_AVX2 AND RUN_AVX2 EQUAL 0)
|
|
|
|
|
set(HAS_AVX2 ON CACHE BOOL "Architecture has AVX2 SIMD enabled")
|
|
|
|
|
message(STATUS "Use AVX2 SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mavx2" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mavx2" )
|
|
|
|
|
add_definitions(-DUSE_AVX2)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /arch:AVX2" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox /arch:AVX2" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_AVX2)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_AVX2 OFF CACHE BOOL "Architecture does not have AVX2 SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_AVX512 COMPILE_AVX512 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx512.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_AVX512 COMPILE_AVX512 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx512.cxx" COMPILE_DEFINITIONS -mavx512f -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_AVX512 AND RUN_AVX512 EQUAL 0)
|
|
|
|
|
set(HAS_AVX512 ON CACHE BOOL "Architecture has AVX512 SIMD enabled")
|
|
|
|
|
message(STATUS "Use AVX512 SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mavx512f" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mavx512f" )
|
|
|
|
|
add_definitions(-DUSE_AVX512)
|
|
|
|
|
elseif(C_MSVC)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /arch:AVX512" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Oi /GL /Ot /Ox /arch:AVX512" )
|
|
|
|
|
set( CMAKE_EXE_LINKER_FLAGS_RELEASE "${CMAKE_EXE_LINKER_FLAGS_RELEASE} /LTCG" )
|
|
|
|
|
add_definitions (/D "_CRT_SECURE_NO_WARNINGS")
|
|
|
|
|
add_definitions(-DUSE_AVX512)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_AVX512 OFF CACHE BOOL "Architecture does not have AVX512 SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
elseif(ARCHITECTURE_ARM)
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
try_run(RUN_NEON COMPILE_NEON "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_arm_neon.cxx" COMPILE_DEFINITIONS /O0)
|
|
|
|
|
else()
|
|
|
|
|
try_run(RUN_NEON COMPILE_NEON "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_arm_neon.cxx" COMPILE_DEFINITIONS -mfpu=neon -O0)
|
|
|
|
|
endif()
|
|
|
|
|
if(COMPILE_NEON AND RUN_NEON EQUAL 0)
|
|
|
|
|
set(HAS_NEON ON CACHE BOOL "Architecture has NEON SIMD enabled")
|
|
|
|
|
message(STATUS "Use NEON SIMD instructions")
|
|
|
|
|
if(C_GCC OR C_CLANG)
|
|
|
|
|
set( CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mfpu=neon" )
|
|
|
|
|
set( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mfpu=neon" )
|
|
|
|
|
add_definitions(-DUSE_NEON)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(HAS_NEON OFF CACHE BOOL "Architecture does not have NEON SIMD enabled")
|
|
|
|
|
endif()
|
|
|
|
|
elseif(ARCHITECTURE_ARM64)
|
|
|
|
|
# Advanced SIMD (aka NEON) is mandatory for AArch64
|
|
|
|
|
set(HAS_NEON ON CACHE BOOL "Architecture has NEON SIMD enabled")
|
|
|
|
|
message(STATUS "Use NEON SIMD instructions")
|
|
|
|
|
add_definitions(-DUSE_NEON)
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
if (FORCE_SSE41)
|
|
|
|
|
message(WARNING "FORCE_SSE41 flag is deprecated, please use ARCH_OPT option.")
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
if (C_MSVC)
|
|
|
|
|
if (ARCHITECTURE_x86)
|
|
|
|
|
set(ARCH_OPT "SSE2")
|
|
|
|
|
else()
|
|
|
|
|
force_ext_available(SSE4_1)
|
|
|
|
|
endif()
|
|
|
|
|
else()
|
|
|
|
|
set(FORCE_OPT "-msse4.1")
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
if (C_MSVC)
|
|
|
|
|
# Glue to make ARCH_OPT more flexible for MSVC
|
|
|
|
|
if (ARCH_OPT STREQUAL "native")
|
|
|
|
|
detect_msvc_native_opt()
|
|
|
|
|
elseif(ARCH_OPT STREQUAL "SSE4_2")
|
|
|
|
|
force_ext_available(SSE4_2)
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
elseif(ARCH_OPT STREQUAL "SSE4_1")
|
|
|
|
|
force_ext_available(SSE4_1)
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
elseif(ARCH_OPT STREQUAL "SSSE3")
|
|
|
|
|
force_ext_available(SSSE3)
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
elseif(ARCH_OPT STREQUAL "SSE3")
|
|
|
|
|
force_ext_available(SSE3)
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
elseif(ARCH_OPT STREQUAL "SSE2")
|
|
|
|
|
force_ext_available(SSE2)
|
|
|
|
|
set(ARCH_OPT "")
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
message(STATUS "Target architecture: ${ARCHITECTURE}-${ARCH_OPT}")
|
|
|
|
|
|
|
|
|
|
cmake_push_check_state(RESET)
|
|
|
|
|
if (ARCH_OPT)
|
|
|
|
|
if(C_MSVC)
|
|
|
|
|
set(CMAKE_REQUIRED_FLAGS "/arch:${ARCH_OPT}")
|
|
|
|
|
add_compile_options(${CMAKE_REQUIRED_FLAGS})
|
|
|
|
|
else()
|
|
|
|
|
set(CMAKE_REQUIRED_FLAGS "-march=${ARCH_OPT}")
|
|
|
|
|
add_compile_options(-march=${ARCH_OPT})
|
|
|
|
|
endif()
|
|
|
|
|
elseif(FORCE_SSSE3 OR FORCE_SSE41)
|
|
|
|
|
if (NOT C_MSVC)
|
|
|
|
|
set(CMAKE_REQUIRED_FLAGS ${FORCE_OPT})
|
|
|
|
|
add_compile_options(${FORCE_OPT})
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
check_cxx_compiler_flag("${CMAKE_REQUIRED_FLAGS}" FLAG_SUPPORTED)
|
|
|
|
|
if (NOT FLAG_SUPPORTED)
|
|
|
|
|
message(FATAL_ERROR "Flag '${CMAKE_REQUIRED_FLAGS}' rejected by compiler. Please adjust ARCH_OPT option.")
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
if (ARCHITECTURE_ARM)
|
|
|
|
|
if (C_MSVC)
|
|
|
|
|
force_ext_available(ARM_NEON)
|
|
|
|
|
else()
|
|
|
|
|
list(APPEND CMAKE_REQUIRED_FLAGS -mfpu=neon)
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
# This is quite basic detection, can be extended if needed
|
|
|
|
|
detect_extensions(ARM_NEON)
|
|
|
|
|
detect_extensions(AVX512F)
|
|
|
|
|
detect_extensions(AVX2 HAS_AVX512F)
|
|
|
|
|
detect_extensions(AVX HAS_AVX2)
|
|
|
|
|
detect_extensions(SSE4_2 HAS_AVX)
|
|
|
|
|
detect_extensions(SSE4_1 HAS_SSE4_2)
|
|
|
|
|
detect_extensions(SSSE3 HAS_SSE4_1)
|
|
|
|
|
detect_extensions(SSE3 HAS_SSSE3)
|
|
|
|
|
detect_extensions(SSE2 HAS_SSE3)
|
|
|
|
|
|
|
|
|
|
cmake_pop_check_state()
|
|
|
|
|
|
|
|
|
|
set(CMAKE_INTERPROCEDURAL_OPTIMIZATION ON)
|
|
|
|
|
if (C_CLANG OR C_GCC)
|
|
|
|
|
add_compile_options(-Wall -Wextra -Wvla -Woverloaded-virtual -ffast-math -ftree-vectorize)
|
|
|
|
|
elseif (C_MSVC)
|
|
|
|
|
add_compile_options(/MP)
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
if (SANITIZE_ADDRESS)
|
|
|
|
|
message(STATUS "Activate address sanitization")
|
|
|
|
|
if(MSVC)
|
|
|
|
|
set(ASAN_LIB_ARCH ${MSVC_CXX_ARCHITECTURE_ID})
|
|
|
|
|
string(TOLOWER ${ASAN_LIB_ARCH} ASAN_LIB_ARCH)
|
|
|
|
|
if(ASAN_LIB_ARCH STREQUAL "x86")
|
|
|
|
|
set(ASAN_LIB_ARCH "i386")
|
|
|
|
|
elseif(ASAN_LIB_ARCH STREQUAL "x64")
|
|
|
|
|
set(ASAN_LIB_ARCH "x86_64")
|
|
|
|
|
endif()
|
|
|
|
|
add_compile_options(/fsanitize=address)
|
|
|
|
|
link_libraries(clang_rt.asan_dynamic-${ASAN_LIB_ARCH} clang_rt.asan_dynamic_runtime_thunk-${ASAN_LIB_ARCH})
|
|
|
|
|
add_link_options(/wholearchive:clang_rt.asan_dynamic_runtime_thunk-${ASAN_LIB_ARCH}.lib)
|
|
|
|
|
else()
|
|
|
|
|
add_compile_options(-fsanitize=address -fno-omit-frame-pointer -g)
|
|
|
|
|
add_link_options(-fsanitize=address)
|
|
|
|
|
endif()
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
# clear binary test folder
|
|
|
|
|