CMake: fix AVX detection on Windows and other SSE/AVX fixes

fvankrieken · Oct 29, 2021 · 350d05e · 350d05e
1 parent c50db0d
commit 350d05e
Show file tree

Hide file tree

Showing 4 changed files with 55 additions and 17 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -137,19 +137,49 @@ if ("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(x86|AMD64)")
   check_compiler_machine_option(flag SSE)
   if (NOT ${flag} EQUAL "")
     set(HAVE_SSE_AT_COMPILE_TIME 1)
-    set(GDAL_SSE_FLAG "-msse")
+    if (NOT ${flag} EQUAL " ")
+      set(GDAL_SSE_FLAG ${flag})
+    endif ()
+  endif ()
+
+  check_compiler_machine_option(flag SSE2)
+  if (NOT ${flag} EQUAL "")
+    set(HAVE_SSE2_AT_COMPILE_TIME 1)
+    if (NOT ${flag} EQUAL " ")
+      set(GDAL_SSE2_FLAG ${flag})
+    endif ()
   endif ()
 
   check_compiler_machine_option(flag SSSE3)
   if (NOT ${flag} EQUAL "")
     set(HAVE_SSSE3_AT_COMPILE_TIME 1)
-    set(GDAL_SSSE3_FLAG "-mssse3")
+    if (NOT ${flag} EQUAL " ")
+      set(GDAL_SSSE3_FLAG ${flag})
+    endif ()
+  endif ()
+
+  check_compiler_machine_option(flag SSE4_1)
+  if (NOT ${flag} EQUAL "")
+    set(HAVE_SSE41_AT_COMPILE_TIME 1)
+    if (NOT ${flag} EQUAL " ")
+      set(GDAL_SSE41_FLAG ${flag})
+    endif ()
   endif ()
 
   check_compiler_machine_option(flag AVX)
   if (NOT ${flag} EQUAL "")
     set(HAVE_AVX_AT_COMPILE_TIME 1)
-    set(GDAL_AVX_FLAG "-mavx")
+    if (NOT ${flag} EQUAL " ")
+      set(GDAL_AVX_FLAG ${flag})
+    endif ()
+  endif ()
+
+  check_compiler_machine_option(flag AVX2)
+  if (NOT ${flag} EQUAL "")
+    set(HAVE_AVX2_AT_COMPILE_TIME 1)
+    if (NOT ${flag} EQUAL " ")
+      set(GDAL_AVX2_FLAG ${flag})
+    endif ()
   endif ()
 
 endif ()

diff --git a/alg/CMakeLists.txt b/alg/CMakeLists.txt
@@ -101,10 +101,16 @@ endif ()
 if (HAVE_AVX_AT_COMPILE_TIME)
   target_sources(alg PRIVATE gdalgridavx.cpp)
   target_compile_definitions(alg PRIVATE -DHAVE_AVX_AT_COMPILE_TIME)
-  set_property(
+  get_property(
+    _compile_flags
     SOURCE gdalgridavx.cpp
-    APPEND
-    PROPERTY COMPILE_FLAGS ${GDAL_AVX_FLAG})
+    PROPERTY COMPILE_FLAGS)
+  if ("${_compile_flags}" STREQUAL "" OR "${_compile_flags}" STREQUAL " ")
+    set(_compile_flags ${GDAL_AVX_FLAG})
+  else ()
+    list(APPEND _compile_flags ${GDAL_AVX_FLAG})
+  endif ()
+  set_property(SOURCE gdalgridavx.cpp PROPERTY COMPILE_FLAGS ${_compile_flags})
 endif ()
 
 include(TargetPublicHeader)

diff --git a/autotest/cpp/CMakeLists.txt b/autotest/cpp/CMakeLists.txt
@@ -186,8 +186,8 @@ register_test(
 if ("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(x86_64|AMD64)")
   gdal_test_target(testsse2 testsse.cpp)
   gdal_test_target(testsse2_emulation testsse.cpp)
+  target_compile_options(testsse2 PRIVATE ${GDAL_SSE2_FLAG})
   target_compile_definitions(testsse2_emulation PRIVATE -DUSE_SSE2_EMULATION -DNO_WARN_USE_SSE2_EMULATION)
-  target_compile_options(testsse2_emulation PRIVATE -msse2)
   register_test(test-sse2 testsse2)
   register_test(test-sse2-emulation testsse2_emulation)
 endif ()
@@ -197,9 +197,9 @@ if (HAVE_AVX_AT_COMPILE_TIME)
   gdal_test_target(testssse3 testsse.cpp)
   gdal_test_target(testsse4_1 testsse.cpp)
   gdal_test_target(testavx2 testsse.cpp)
-  target_compile_options(testssse3 PRIVATE -mssse3)
-  target_compile_options(testsse4_1 PRIVATE -mssse3 -msse4.1)
-  target_compile_options(testavx2 PRIVATE -mavx2)
+  target_compile_options(testssse3 PRIVATE ${GDAL_SSSE3_FLAG})
+  target_compile_options(testsse4_1 PRIVATE ${GDAL_SSSE3_FLAG} ${GDAL_SSSE41_FLAG})
+  target_compile_options(testavx2 PRIVATE ${GDAL_AVX2_FLAG})
   register_test_as_custom_target(test-ssse3 testssse3)
   register_test_as_custom_target(test-sse41 testsse4_1)
   register_test_as_custom_target(test-avx2 testavx2)

diff --git a/cmake/helpers/CheckCompilerMachineOption.cmake b/cmake/helpers/CheckCompilerMachineOption.cmake
@@ -22,27 +22,30 @@ include(CheckIncludeFileCXX)
 include(CheckIncludeFile)
 
 function(check_compiler_machine_option outvar feature)
- macro(__check_compiler_flag _flag _result)
+  macro(__check_compiler_flag _flag _result)
     if(CMAKE_CXX_COMPILER_LOADED)
       check_cxx_compiler_flag("${_flag}" ${_result})
     elseif(CMAKE_C_COMPILER_LOADED)
       check_c_compiler_flag("${_flag}" ${_result})
     endif()
   endmacro()
 
+  set(_FLAGS)
   if("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(x86|AMD64)")
-    if(MSVC)
+    if(MSVC AND (${feature} MATCHES "SSE"))
+      # SSE2 and SSE are default on
+	  set(_FLAGS " ")
+    elseif(MSVC)
       # Only Visual Studio 2017 version 15.3 / Visual C++ 19.11 & up have support for AVX-512.
       # https://blogs.msdn.microsoft.com/vcblog/2017/07/11/microsoft-visual-studio-2017-supports-intel-avx-512/
-      # SSE2 and SSE are default on
       set(MSVC_FLAG_MAP AVX512 "-arch:AVX512" AVX2 "-arch:AVX2" AVX "-arch:AVX")
       list(FIND MSVC_FLAG_MAP ${feature} _found)
       if(_found GREATER -1)
         math(EXPR index "${_found}+1")
         list(GET MSVC_FLAG_MAP ${index} _flag)
         __check_compiler_flag("${_flag}" test_${feature})
         if(test_${feature})
-          set(_FLAGS "${flag}")
+          set(_FLAGS "${_flag}")
         endif()
       endif()
     elseif(CMAKE_CXX_COMPILER MATCHES "/(icpc|icc)$") # ICC (on Linux)
@@ -62,6 +65,7 @@ function(check_compiler_machine_option outvar feature)
       endif()
     else() # not MSVC and not ICC => GCC, Clang, Open64
       string(TOLOWER ${feature} _flag)
+	  string(REPLACE "_" "." _flag "${_flag}")
       __check_compiler_flag("-m${_flag}" test_${_flag})
       if(test_${_flag})
         set(header_table "sse3" "pmmintrin.h" "ssse3" "tmmintrin.h" "sse4.1" "smmintrin.h"
@@ -87,9 +91,7 @@ function(check_compiler_machine_option outvar feature)
         endif()
       endif()
     endif()
-    set(${outvar} ${_FLAGS} PARENT_SCOPE)
   elseif("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(ARM|aarch64)")
-    set(_FLAGS)
     if(MSVC)
       # TODO implement me
     elseif(CMAKE_CXX_COMPILER MATCHES "/(icpc|icc)$") # ICC (on Linux)
@@ -101,6 +103,6 @@ function(check_compiler_machine_option outvar feature)
         set(_FLAGS "-m${_flag}")
       endif()
     endif()
-    set(${outvar} "${_FLAGS}" PARENT_SCOPE)
   endif()
+  set(${outvar} "${_FLAGS}" PARENT_SCOPE)
 endfunction()