Add support for Windows ARM64

Signed-off-by: Anthony Roberts <[email protected]>

Author: anthony-linaro

CMakeLists.txt

@@ -270,7 +270,7 @@ option(OCIO_USE_AVX2 "Specify whether to enable AVX2 CPU performance optimizatio
 option(OCIO_USE_AVX512 "Specify whether to enable AVX512 CPU performance optimizations" ${OCIO_BUILD_ENABLE_OPTIMIZATIONS_AVX})
 option(OCIO_USE_F16C  "Specify whether to enable F16C CPU performance optimizations" ${OCIO_BUILD_ENABLE_OPTIMIZATIONS_F16C})
 
-if (APPLE)
+if (APPLE OR WIN32)
    # TODO: Revisit whether that option is necessary.
    option(OCIO_USE_SSE2NEON "Specify whether to enable SSE CPU performance optimizations using SSE2NEON for Apple ARM architecture" ON)
    mark_as_advanced(OCIO_USE_SSE2NEON)
@@ -332,8 +332,10 @@ if(OCIO_USE_SIMD AND OCIO_USE_SSE2NEON AND COMPILER_SUPPORTS_ARM_NEON)
         add_library(sse2neon INTERFACE)
         # Add the include directories to the target.
         target_include_directories(sse2neon INTERFACE "${sse2neon_INCLUDE_DIR}")
-        # Ignore the warnings coming from sse2neon.h as they are false positives.
-        target_compile_options(sse2neon INTERFACE -Wno-unused-parameter)
+        if(NOT MSVC)
+            # Ignore the warnings coming from sse2neon.h as they are false positives.
+            target_compile_options(sse2neon INTERFACE -Wno-unused-parameter)
+        endif()
     endif()
 endif()
 

share/cmake/modules/install/Installsse2neon.cmake

@@ -16,7 +16,7 @@ include(FetchContent)
 set(FETCHCONTENT_BASE_DIR "${CMAKE_BINARY_DIR}/ext/build/sse2neon")
 FetchContent_Declare(sse2neon
   GIT_REPOSITORY https://github.com/DLTcollab/sse2neon.git
-  GIT_TAG        v1.6.0
+  GIT_TAG        227cc413fb2d50b2a10073087be96b59d5364aea
 )
 
 # FetchContent_MakeAvailable is not available until CMake 3.14+.
@@ -38,6 +38,8 @@ if(NOT sse2neon_POPULATED)
   add_library(sse2neon INTERFACE)
   # Add the include directories to the target.
   target_include_directories(sse2neon INTERFACE "${sse2neon_INCLUDE_DIR}")
-  # Ignore the warnings coming from sse2neon.h as they are false positives.
-  target_compile_options(sse2neon INTERFACE -Wno-unused-parameter)
+  if(NOT MSVC)
+      # Ignore the warnings coming from sse2neon.h as they are false positives.
+      target_compile_options(sse2neon INTERFACE -Wno-unused-parameter)
+  endif()
 endif()

share/cmake/utils/CheckSupportSSEUsingSSE2NEON.cmake

@@ -6,8 +6,13 @@ include(CheckCXXSourceCompiles)
 set(_cmake_required_flags_orig "${CMAKE_REQUIRED_FLAGS}")
 set(_cmake_required_includes_orig "${CMAKE_REQUIRED_INCLUDES}")
 set(_cmake_osx_architectures_orig "${CMAKE_OSX_ARCHITECTURES}")
+set(_cmake_cxx_flags_orig "${CMAKE_CXX_FLAGS}")
 
-if(APPLE AND COMPILER_SUPPORTS_ARM_NEON)
+if(MSVC)
+    set(CMAKE_CXX_FLAGS "/Zc:preprocessor")
+endif()
+
+if((APPLE OR WIN32) AND COMPILER_SUPPORTS_ARM_NEON)
 
     if("${CMAKE_OSX_ARCHITECTURES}" MATCHES "arm64;x86_64" OR
        "${CMAKE_OSX_ARCHITECTURES}" MATCHES "x86_64;arm64")
@@ -63,8 +68,9 @@ endif()
 set(CMAKE_REQUIRED_FLAGS "${_cmake_required_flags_orig}")
 set(CMAKE_REQUIRED_INCLUDES "${_cmake_required_includes_orig}")
 set(CMAKE_OSX_ARCHITECTURES "${_cmake_osx_architectures_orig}")
+set(CMAKE_CXX_FLAGS "${_cmake_cxx_flags_orig}")
 
 unset(_cmake_required_flags_orig)
 unset(_cmake_required_includes_orig)
 unset(_cmake_osx_architectures_orig)
-
+unset(_cmake_cxx_flags_orig)

share/cmake/utils/CompilerFlags.cmake

@@ -19,6 +19,9 @@ if(OCIO_USE_SIMD)
     if (OCIO_USE_SSE2NEON AND COMPILER_SUPPORTS_ARM_NEON)
         include(CheckSupportSSEUsingSSE2NEON)
         if(NOT COMPILER_SUPPORTS_SSE_WITH_SSE2NEON)
+            # Enable the "new" preprocessor, to more closely match Clang/GCC, required for sse2neon
+            set(PLATFORM_COMPILE_OPTIONS "${PLATFORM_COMPILE_OPTIONS};/Zc:preprocessor")
+        else()
             set(OCIO_USE_SSE2NEON OFF)
         endif()
     endif()

src/OpenColorIO/CPUInfo.cpp

@@ -183,7 +183,7 @@ CPUInfo::CPUInfo()
     }
 }
 
-#elif defined(__aarch64__) // ARM Processor or Apple ARM.
+#elif defined(__aarch64__) || defined(_M_ARM64) // ARM Processor or Apple ARM.
 
 CPUInfo::CPUInfo()
 {

src/OpenColorIO/CPUInfoConfig.h.in

@@ -6,7 +6,7 @@
 #cmakedefine01 OCIO_ARCH_X86_32
 
 // Relevant only for arm64 architecture.
-#if defined(__aarch64__)
+#if defined(__aarch64__) || defined(_M_ARM64)
     #cmakedefine01 OCIO_USE_SSE2NEON
 #else
     #define OCIO_USE_SSE2NEON 0
@@ -23,7 +23,7 @@
 
 // Building for x86_64 processor on a non-ARM host architecture
 // OR Building on/for an ARM architecture and using SSE2NEON.
-#if (OCIO_ARCH_X86 && !defined(__aarch64__)) || (defined(__aarch64__) && OCIO_USE_SSE2NEON)
+#if (OCIO_ARCH_X86 && !defined(__aarch64__)) || ((defined(__aarch64__) || defined(_M_ARM64)) && OCIO_USE_SSE2NEON)
     #cmakedefine01 OCIO_USE_SSE2
     #cmakedefine01 OCIO_USE_SSE3
     #cmakedefine01 OCIO_USE_SSSE3

src/OpenColorIO/SSE.h

@@ -9,11 +9,11 @@
 #if OCIO_USE_SSE2
 
 // Include the appropriate SIMD intrinsics header based on the architecture (Intel vs. ARM).
-#if !defined(__aarch64__)
+#if !defined(__aarch64__) && !defined(_M_ARM64)
     #if OCIO_USE_SSE2
         #include <emmintrin.h>
     #endif
-#elif defined(__aarch64__)
+#elif defined(__aarch64__) || defined(_M_ARM64)
     // ARM architecture A64 (ARM64)
     #if OCIO_USE_SSE2NEON
         #include <sse2neon.h>
@@ -30,7 +30,7 @@ namespace OCIO_NAMESPACE
 // Note that it is important for the code below this ifdef stays in the OCIO_NAMESPACE since
 // it is redefining two of the functions from sse2neon.
 
-#if defined(__aarch64__)
+#if defined(__aarch64__) || defined(_M_ARM64)
     #if OCIO_USE_SSE2NEON
         // Using vmaxnmq_f32 and vminnmq_f32 rather than sse2neon's vmaxq_f32 and vminq_f32 due to 
         // NaN handling. This doesn't seem to be significantly slower than the default sse2neon behavior.

src/OpenColorIO/SSE2.h

@@ -9,9 +9,9 @@
 #if OCIO_USE_SSE2
 
 // Include the appropriate SIMD intrinsics header based on the architecture (Intel vs. ARM).
-#if !defined(__aarch64__)
+#if !defined(__aarch64__) && !defined(_M_ARM64)
     #include <emmintrin.h>
-#elif defined(__aarch64__)
+#elif defined(__aarch64__) || defined(_M_ARM64)
     // ARM architecture A64 (ARM64)
     #if OCIO_USE_SSE2NEON
         #include <sse2neon.h>
@@ -30,7 +30,7 @@ namespace OCIO_NAMESPACE
 // Note that it is important for the code below this ifdef stays in the OCIO_NAMESPACE since
 // it is redefining two of the functions from sse2neon.
 
-#if defined(__aarch64__)
+#if defined(__aarch64__) || defined(_M_ARM64)
     #if OCIO_USE_SSE2NEON
         // Using vmaxnmq_f32 and vminnmq_f32 rather than sse2neon's vmaxq_f32 and vminq_f32 due to 
         // NaN handling. This doesn't seem to be significantly slower than the default sse2neon behavior.