Use SSE4.1 on Windows

Author: jsiegle

CMakeLists.txt

@@ -106,7 +106,8 @@ if(BUILD_TESTS)
 	add_library(gui_testable_source SHARED ${JUCE_SOURCE})
 	if(MSVC)
 		set_property(TARGET gui_testable_source APPEND PROPERTY COMPILE_DEFINITIONS "JUCE_API=__declspec(dllexport)")
-		target_compile_options(gui_testable_source PRIVATE /sdl- /nologo /MP /W0 /bigobj)
+		# /arch:AVX enables SSE4.1 intrinsics (required for proper SIMD rounding in SIMDConverter)
+		target_compile_options(gui_testable_source PRIVATE /sdl- /nologo /MP /W0 /bigobj /arch:AVX)
 	endif()
 	target_include_directories(gui_testable_source PRIVATE ${JUCE_DIRECTORY} ${JUCE_DIRECTORY}/modules)
 	target_compile_features(gui_testable_source PUBLIC cxx_auto_type cxx_generalized_initializers cxx_std_17)
@@ -188,7 +189,8 @@ endif()
 
 if(MSVC)
 	set_property(TARGET open-ephys APPEND PROPERTY COMPILE_DEFINITIONS "JUCE_API=__declspec(dllexport)")
-	target_compile_options(open-ephys PRIVATE /sdl- /nologo /MP /W0 /bigobj)
+	# /arch:AVX enables SSE4.1 intrinsics (required for proper SIMD rounding in SIMDConverter)
+	target_compile_options(open-ephys PRIVATE /sdl- /nologo /MP /W0 /bigobj /arch:AVX)
 	target_link_libraries(open-ephys setupapi.lib opengl32.lib glu32.lib)
 	set_property(TARGET open-ephys APPEND PROPERTY VS_DEBUGGER_COMMAND "Debug\\open-ephys.exe")
 	set_property(TARGET open-ephys APPEND_STRING PROPERTY LINK_FLAGS_DEBUG " /NODEFAULTLIB:\"libcmt.lib\" /NODEFAULTLIB:\"msvcrt.lib\"")

Source/Processors/RecordNode/BinaryFormat/BinaryRecording.cpp

@@ -40,6 +40,8 @@ BinaryRecording::BinaryRecording()
     m_scaledBuffer.malloc (MAX_BUFFER_SIZE);
     m_intBuffer.malloc (MAX_BUFFER_SIZE);
     m_sampleNumberBuffer.malloc (MAX_BUFFER_SIZE);
+
+    LOGD ("BinaryRecording SIMD type: ", SIMDConverter::getSIMDTypeString());
 }
 
 BinaryRecording::~BinaryRecording() {}

Source/Processors/RecordNode/BinaryFormat/SIMDConverter.cpp

@@ -273,8 +273,7 @@ void SIMDConverter::convertSSE4_1 (const float* input, int16_t* output, float sc
         f0 = _mm_mul_ps (f0, vscale);
         f1 = _mm_mul_ps (f1, vscale);
 
-        // Round to nearest integer (SSE4.1)
-        // _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC = 0x00 | 0x08 = 0x08
+        // Round to nearest integer (SSE4.1 - uses banker's rounding)
         f0 = _mm_round_ps (f0, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
         f1 = _mm_round_ps (f1, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
 

Source/Processors/RecordNode/BinaryFormat/SIMDConverter.h

@@ -24,6 +24,11 @@
 #ifndef SIMDCONVERTER_H
 #define SIMDCONVERTER_H
 
+// MSVC with /arch:AVX doesn't define __SSE4_1__, but AVX implies SSE4.1 support
+#if defined(_MSC_VER) && defined(__AVX__) && !defined(__SSE4_1__)
+    #define __SSE4_1__ 1
+#endif
+
 #include <cstdint>
 #include <string>