ref: 84484a679009fc703d80cedfafebb50d42b97cd4
parent: e9dc5d1793d8add52a6a70a0c286370adac50d95
author: Marcus Asteborg <xnorpx@outlook.com>
date: Wed May 24 17:28:36 EDT 2023
CMake - change avx to avx2 and fma
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -177,13 +177,13 @@
OFF)
add_feature_info(OPUS_X86_MAY_HAVE_SSE4_1 OPUS_X86_MAY_HAVE_SSE4_1 ${OPUS_X86_MAY_HAVE_SSE4_1_HELP_STR})- set(OPUS_X86_MAY_HAVE_AVX_HELP_STR "does runtime check for AVX support.")
- cmake_dependent_option(OPUS_X86_MAY_HAVE_AVX
- ${OPUS_X86_MAY_HAVE_AVX_HELP_STR}+ set(OPUS_X86_MAY_HAVE_AVX2_HELP_STR "does runtime check for AVX FMA AVX2 support.")
+ cmake_dependent_option(OPUS_X86_MAY_HAVE_AVX2
+ ${OPUS_X86_MAY_HAVE_AVX2_HELP_STR}ON
- "AVX_SUPPORTED; NOT OPUS_DISABLE_INTRINSICS"
+ "AVX2_SUPPORTED; NOT OPUS_DISABLE_INTRINSICS"
OFF)
- add_feature_info(OPUS_X86_MAY_HAVE_AVX OPUS_X86_MAY_HAVE_AVX ${OPUS_X86_MAY_HAVE_AVX_HELP_STR})+ add_feature_info(OPUS_X86_MAY_HAVE_AVX2 OPUS_X86_MAY_HAVE_AVX2 ${OPUS_X86_MAY_HAVE_AVX2_HELP_STR})# PRESUME depends on MAY HAVE, but PRESUME will override runtime detection
set(OPUS_X86_PRESUME_SSE_HELP_STR "assume target CPU has SSE1 support (override runtime check).")
@@ -224,13 +224,13 @@
OFF)
add_feature_info(OPUS_X86_PRESUME_SSE4_1 OPUS_X86_PRESUME_SSE4_1 ${OPUS_X86_PRESUME_SSE4_1_HELP_STR})- set(OPUS_X86_PRESUME_AVX_HELP_STR "assume target CPU has AVX support (override runtime check).")
- cmake_dependent_option(OPUS_X86_PRESUME_AVX
- ${OPUS_X86_PRESUME_AVX_HELP_STR}+ set(OPUS_X86_PRESUME_AVX2_HELP_STR "assume target CPU has AVX FMA AVX2 support (override runtime check).")
+ cmake_dependent_option(OPUS_X86_PRESUME_AVX2
+ ${OPUS_X86_PRESUME_AVX2_HELP_STR}OFF
- "OPUS_X86_MAY_HAVE_AVX; NOT OPUS_DISABLE_INTRINSICS"
+ "OPUS_X86_MAY_HAVE_AVX2; NOT OPUS_DISABLE_INTRINSICS"
OFF)
- add_feature_info(OPUS_X86_PRESUME_AVX OPUS_X86_PRESUME_AVX ${OPUS_X86_PRESUME_AVX_HELP_STR})+ add_feature_info(OPUS_X86_PRESUME_AVX2 OPUS_X86_PRESUME_AVX2 ${OPUS_X86_PRESUME_AVX2_HELP_STR})endif()
feature_summary(WHAT ALL)
@@ -382,7 +382,7 @@
if(((OPUS_X86_MAY_HAVE_SSE AND NOT OPUS_X86_PRESUME_SSE) OR
(OPUS_X86_MAY_HAVE_SSE2 AND NOT OPUS_X86_PRESUME_SSE2) OR
(OPUS_X86_MAY_HAVE_SSE4_1 AND NOT OPUS_X86_PRESUME_SSE4_1) OR
- (OPUS_X86_MAY_HAVE_AVX AND NOT OPUS_X86_PRESUME_AVX)) AND
+ (OPUS_X86_MAY_HAVE_AVX2 AND NOT OPUS_X86_PRESUME_AVX2)) AND
RUNTIME_CPU_CAPABILITY_DETECTION)
target_compile_definitions(opus PRIVATE OPUS_HAVE_RTCD)
if(NOT MSVC)
@@ -454,22 +454,28 @@
endif()
endif()
- if(AVX_SUPPORTED)
- # mostly placeholder in case of avx intrinsics is added
- if(OPUS_X86_MAY_HAVE_AVX)
- target_compile_definitions(opus PRIVATE OPUS_X86_MAY_HAVE_AVX)
+ if(AVX2_SUPPORTED)
+ if(OPUS_X86_MAY_HAVE_AVX2)
+ add_sources_group(opus celt ${celt_sources_avx2})+ target_compile_definitions(opus PRIVATE OPUS_X86_MAY_HAVE_AVX2)
+ if(MSVC)
+ set(AVX2_FLAGS "${AVX2_FLAGS} /arch:AVX2")+ else()
+ set(AVX2_FLAGS "${AVX2_FLAGS} -mavx2 -mfma -mavx")+ endif()
+ set_source_files_properties(${celt_sources_avx2} PROPERTIES COMPILE_FLAGS ${AVX2_FLAGS})endif()
- if(OPUS_X86_PRESUME_AVX)
- target_compile_definitions(opus PRIVATE OPUS_X86_PRESUME_AVX)
+ if(OPUS_X86_PRESUME_AVX2)
+ target_compile_definitions(opus PRIVATE OPUS_X86_PRESUME_AVX2)
if(NOT MSVC)
- target_compile_options(opus PRIVATE -mavx)
+ target_compile_options(opus PRIVATE -mavx2 -mfma -mavx)
endif()
endif()
endif()
if(MSVC)
- if(AVX_SUPPORTED AND OPUS_X86_PRESUME_AVX) # on 64 bit and 32 bits
- add_definitions(/arch:AVX)
+ if(AVX2_SUPPORTED AND OPUS_X86_PRESUME_AVX2) # on 64 bit and 32 bits
+ add_definitions(/arch:AVX2)
elseif(OPUS_CPU_X86) # if AVX not supported then set SSE flag
if((SSE4_1_SUPPORTED AND OPUS_X86_PRESUME_SSE4_1)
OR (SSE2_SUPPORTED AND OPUS_X86_PRESUME_SSE2))
@@ -602,10 +608,6 @@
add_executable(opus_compare ${opus_compare_sources}) target_include_directories(opus_compare PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) target_link_libraries(opus_compare PRIVATE opus ${OPUS_REQUIRED_LIBRARIES})- if(MSVC)
- # move cosmetic warning to level 4 for opus_compare
- target_compile_options(opus_compare PRIVATE /w44244)
- endif()
endif()
if(BUILD_TESTING AND NOT BUILD_SHARED_LIBS)
--- a/cmake/OpusConfig.cmake
+++ b/cmake/OpusConfig.cmake
@@ -102,7 +102,10 @@
endif()
endif()
-if(NOT MSVC)
+if(MSVC)
+ # move cosmetic warnings to level 4
+ add_compile_options(/w44244 /w44305 /w44267)
+else()
set(WARNING_LIST -Wall -W -Wstrict-prototypes -Wextra -Wcast-align -Wnested-externs -Wshadow)
include(CheckCCompilerFlag)
foreach(WARNING_FLAG ${WARNING_LIST})--- a/cmake/OpusFunctions.cmake
+++ b/cmake/OpusFunctions.cmake
@@ -47,10 +47,12 @@
endfunction()
include(CheckIncludeFile)
-# function to check if compiler supports SSE, SSE2, SSE4.1 and AVX if target
-# systems may not have SSE support then use OPUS_MAY_HAVE_SSE option if target
-# system is guaranteed to have SSE support then OPUS_PRESUME_SSE can be used to
-# skip SSE runtime check
+
+# This function determines if the compiler has support for SSE, SSE2, SSE4.1, AVX,
+# AVX2 and FMA. Should the target systems potentially lack SSE support, the
+# OPUS_MAY_HAVE_SSE option is recommended for use. If, however, the target system is
+# assured to support SSE, the OPUS_PRESUME_SSE option can be employed, thus
+# eliminating the necessity for an SSE runtime check.
function(opus_detect_sse COMPILER_SUPPORT_SIMD)
message(STATUS "Check SIMD support by compiler")
check_include_file(xmmintrin.h HAVE_XMMINTRIN_H) # SSE1
@@ -111,20 +113,20 @@
PARENT_SCOPE)
endif()
- check_include_file(immintrin.h HAVE_IMMINTRIN_H) # AVX
+ check_include_file(immintrin.h HAVE_IMMINTRIN_H) # AVX2
if(HAVE_IMMINTRIN_H)
if(MSVC)
- check_flag(AVX /arch:AVX)
+ check_flag(AVX2 /arch:AVX2)
else()
- check_flag(AVX -mavx)
+ check_flag(AVX2 -mavx2 -mfma -mavx)
endif()
else()
- set(AVX_SUPPORTED
+ set(AVX2_SUPPORTED
0
PARENT_SCOPE)
endif()
- if(SSE1_SUPPORTED OR SSE2_SUPPORTED OR SSE4_1_SUPPORTED OR AVX_SUPPORTED)
+ if(SSE1_SUPPORTED OR SSE2_SUPPORTED OR SSE4_1_SUPPORTED OR AVX2_SUPPORTED)
set(COMPILER_SUPPORT_SIMD 1 PARENT_SCOPE)
else()
message(STATUS "No SIMD support in compiler")
--- a/cmake/OpusSources.cmake
+++ b/cmake/OpusSources.cmake
@@ -29,6 +29,7 @@
get_opus_sources(CELT_SOURCES_SSE celt_sources.mk celt_sources_sse)
get_opus_sources(CELT_SOURCES_SSE2 celt_sources.mk celt_sources_sse2)
get_opus_sources(CELT_SOURCES_SSE4_1 celt_sources.mk celt_sources_sse4_1)
+get_opus_sources(CELT_SOURCES_AVX2 celt_sources.mk celt_sources_avx2)
get_opus_sources(CELT_SOURCES_ARM_RTCD celt_sources.mk celt_sources_arm_rtcd)
get_opus_sources(CELT_SOURCES_ARM_ASM celt_sources.mk celt_sources_arm_asm)
get_opus_sources(CELT_AM_SOURCES_ARM_ASM celt_sources.mk
--
⑨