Add support for Windows ARM64 (#2089)

* Add support for Windows ARM64

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

* Fix improper compiler flag check

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

* Fix sse2neon issues on Windows ARM64

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

* Fix cross-compilation on Windows for X64 -> ARM64

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

* Fix comment to match with corresponding if directive

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

* Check for MSVC before setting MSVC-style flag

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

* Fix comment to resolve ambiguity

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>

---------

Signed-off-by: Anthony Roberts <anthony.roberts@linaro.org>
Co-authored-by: Doug Walker <doug.walker@autodesk.com>

Author: anthony-linaro

CMakeLists.txt

@@ -180,7 +180,12 @@ option(OCIO_USE_OIIO_FOR_APPS "Request OIIO to build apps (ociolutimage, ociocon
 
 
 if (NOT APPLE)
-    if ("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(AMD64|IA64|EM64T|x86_64|X86|i386|i686)")
+    if("${CMAKE_GENERATOR_PLATFORM}" MATCHES "(ARM64|arm64)" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "ARM64")
+        set(OCIO_ARCH_X86 0)
+        set(OCIO_BUILD_ENABLE_OPTIMIZATIONS_SSE ON)
+        set(OCIO_BUILD_ENABLE_OPTIMIZATIONS_AVX OFF)
+        set(OCIO_BUILD_ENABLE_OPTIMIZATIONS_F16C OFF)
+    elseif ("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(AMD64|IA64|EM64T|x86_64|X86|i386|i686)")
         # Intel-based architecture (not APPLE)
         if ("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "(X86|i386|i686)")
             set(OCIO_ARCH_X86_32 1)
@@ -270,7 +275,7 @@ 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 +337,10 @@ 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

@@ -18,7 +18,12 @@ if(OCIO_USE_SIMD)
 
     if (OCIO_USE_SSE2NEON AND COMPILER_SUPPORTS_ARM_NEON)
         include(CheckSupportSSEUsingSSE2NEON)
-        if(NOT COMPILER_SUPPORTS_SSE_WITH_SSE2NEON)
+        if(COMPILER_SUPPORTS_SSE_WITH_SSE2NEON)
+            if(WIN32 AND MSVC)
+                # Enable the "new" preprocessor, to more closely match Clang/GCC, required for sse2neon
+                set(PLATFORM_COMPILE_OPTIONS "${PLATFORM_COMPILE_OPTIONS};/Zc:preprocessor")
+            endif()
+        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 64-bit processor (multiple platforms)
 
 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

@@ -9,14 +9,25 @@
 #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
+        // MSVC doesn't like the redefinitions below and requires the existing functions to be undef-ed
+        #if defined(_M_ARM64)
+            #define _mm_max_ps _mm_max_ps_orig
+            #define _mm_min_ps _mm_min_ps_orig
+        #endif
+
         #include <sse2neon.h>
+
+        #if defined(_M_ARM64)
+            #undef _mm_max_ps
+            #undef _mm_min_ps
+        #endif
     #endif
 #endif
 
@@ -30,7 +41,7 @@ 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.
@@ -77,6 +88,9 @@ static const __m128 EPOS128 = _mm_set1_ps(128.0f);
 
 static const __m128 EPOSINF = _mm_set1_ps(std::numeric_limits<float>::infinity());
 
+// These funtions won't work when using MSVC + ARM64 unless you specify /Zc:arm64-aliased-neon-types-
+// This comes with it's own issues, so it is easier to just disable them when using MSVC + ARM64
+#if !defined(_M_ARM64)
 // Debug function to print out the contents of a floating-point SSE register
 inline void ssePrintRegister(const char* msg, __m128& reg)
 {
@@ -91,6 +105,7 @@ inline void ssePrintRegister(const char* msg, __m128i& reg)
     int *r = (int*) &reg;
     printf("%s : %d %d %d %d\n", msg, r[0], r[1], r[2], r[3]);
 }
+#endif
 
 // Determine whether a floating-point value is negative based on its sign bit.
 // This function will treat special values, like -0, -NaN, -Inf, as they were indeed
@@ -170,7 +185,7 @@ inline __m128 sseLog2(__m128 x)
 {
     // y = log2( x ) = log2( 2^exponent * mantissa ) 
     //               = exponent + log2( mantissa )
-
+  
     __m128 mantissa
         = _mm_or_ps(                                    // OR with EONE
             _mm_andnot_ps(                              // NOT(EMASK) AND x

src/OpenColorIO/SSE2.h

@@ -9,12 +9,35 @@
 #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
+        // MSVC doesn't like the redefinitions below and requires the existing functions to be undef-ed
+        #if defined(_M_ARM64)
+            #define _mm_max_ps _mm_max_ps_orig
+            #define _mm_min_ps _mm_min_ps_orig
+        #endif
+
         #include <sse2neon.h>
+
+        #if defined(_M_ARM64)
+            #undef _mm_max_ps
+            #undef _mm_min_ps
+        #endif
+
+        // Current versions of MSVC do not define float16_t, so we do it ourselves using
+        // int16_t as an intermediate type
+        #if defined(_M_ARM64) && !defined(float16_t)
+            #define float16_t int16_t
+        #endif
+
+        // Current versions of MSVC do not define vst1q_f16, so we do it ourselves using
+        // internal methods from MSVC's arm_neon.h
+        #if defined(_M_ARM64) && !defined(vst1q_f16)
+            #define vst1q_f16(A, B) neon_st1m_q16((A), __float16x8_t_to_n128(B));
+        #endif
     #endif
 #endif
 
@@ -30,7 +53,7 @@ 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.
@@ -321,7 +344,7 @@ struct SSE2RGBAPack<BIT_DEPTH_F16>
         sse2RGBATranspose_4x4(r, g, b, a, rgba0, rgba1, rgba2, rgba3);
 
 #if OCIO_USE_SSE2NEON
-        // use neon hardware support for f32 to f16
+        // use neon hardware support for f32 to f16 (apart from in MSVC, which doesnt support it)
         float16x8_t rgba;
         float16x4_t rgba00_01 = vcvt_f16_f32(vreinterpretq_f32_m128(rgba0));
         float16x4_t rgba03_03 = vcvt_f16_f32(vreinterpretq_f32_m128(rgba1));

src/OpenColorIO/ops/fixedfunction/FixedFunctionOpCPU.cpp

@@ -1844,7 +1844,7 @@ __m128 Renderer_LIN_TO_PQ_SSE<true>::myPower(__m128 x, __m128 exp)
     return ssePower(x, exp);
 }
 
-#ifdef _WIN32
+#if (_MSC_VER >= 1920) && (OCIO_USE_AVX)
 // Only Windows compilers have built-in _mm_pow_ps() SVML intrinsic
 // implementation, so non-fast SIMD version is available only on Windows for
 // now.
@@ -1853,7 +1853,7 @@ __m128 Renderer_LIN_TO_PQ_SSE<false>::myPower(__m128 x, __m128 exp)
 {
     return _mm_pow_ps(x, exp);
 }
-#endif // _WIN32
+#endif // (_MSC_VER >= 1920) && (OCIO_USE_AVX)
 
 template<bool FAST_POWER>
 void Renderer_LIN_TO_PQ_SSE<FAST_POWER>::apply(const void* inImg, void* outImg, long numPixels) const