GPU:

- Avoid generating autovectorized SSE2 code for loops where a hand-coded SSE2 loop already exists. (MSVC and Clang only.)
2016-06-23 20:30:24 +00:00 · 2016-06-23 20:30:24 +00:00 · dde0da24ab
parent 3f895b85fb
commit dde0da24ab
6 changed files with 120 additions and 36 deletions
--- a/desmume/src/GPU.cpp
+++ b/desmume/src/GPU.cpp
@ -2917,6 +2917,10 @@ void GPUEngineBase::_RenderPixelsCustom(void *__restrict dstColorLine, u8 *__res
 																												   (GPU->GetDisplayInfo().colorFormat == NDSColorFormat_BGR555_Rev) ? (u16 *)(dstColorLine16 + i) : (u16 *)(dstColorLine32 + i),
 																												   dstLayerID + i);
 		}
+#endif
+		
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 		for (; i < dstPixCount; i++)
 		{
@ -2969,6 +2973,10 @@ void GPUEngineBase::_RenderPixelsCustomVRAM(void *__restrict dstColorLine, u8 *_
 																											   (GPU->GetDisplayInfo().colorFormat == NDSColorFormat_BGR555_Rev) ? (u16 *)(dstColorLine16 + i) : (u16 *)(dstColorLine32 + i),
 																											   dstLayerID + i);
 	}
+#endif
+	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 	for (; i < dstPixCount; i++)
 	{
@ -4004,6 +4012,10 @@ void GPUEngineBase::ApplyMasterBrightness()
 							const __m128i dstColor_vec128 = _mm_load_si128((__m128i *)((u16 *)dst + i));
 							_mm_store_si128( (__m128i *)((u16 *)dst + i), this->_ColorEffectIncreaseBrightness(dstColor_vec128, intensity_vec128) );
 						}
+#endif
+						
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 						for (; i < pixCount; i++)
 						{
@ -4077,6 +4089,10 @@ void GPUEngineBase::ApplyMasterBrightness()
 							const __m128i dstColor_vec128 = _mm_load_si128((__m128i *)((u16 *)dst + i));
 							_mm_store_si128( (__m128i *)((u16 *)dst + i), this->_ColorEffectDecreaseBrightness(dstColor_vec128, intensity_vec128) );
 						}
+#endif
+						
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 						for (; i < pixCount; i++)
 						{
@ -5041,6 +5057,10 @@ void* GPUEngineA::_RenderLine_Layers(const u16 l)
 																	(dispInfo.colorFormat == NDSColorFormat_BGR555_Rev) ? (u16 *)(dstColorLine16 + dstX) : (u16 *)(dstColorLine32 + dstX),
 																	dstLayerIDPtr + dstX);
 								}
+#endif
+								
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 								for (; dstX < customLineWidth; dstX++)
 								{
@ -5669,6 +5689,10 @@ void GPUEngineA::_RenderLine_DispCapture_Copy(const u16 *src, u16 *dst, const si
 				{
 					_mm_store_si128((__m128i *)(dst + i), _mm_or_si128( _mm_load_si128( (__m128i *)(src + i)), alpha_vec128 ) );
 				}
+#endif
+				
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 				for (; i < pixCountExt; i++)
 				{
@ -5686,6 +5710,10 @@ void GPUEngineA::_RenderLine_DispCapture_Copy(const u16 *src, u16 *dst, const si
 					{
 						_mm_store_si128((__m128i *)(dst + i), _mm_or_si128( _mm_load_si128( (__m128i *)(src + i)), alpha_vec128 ) );
 					}
+#endif
+					
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 					for (; i < captureLengthExt; i++)
 					{
@ -5824,6 +5852,9 @@ void GPUEngineA::_RenderLine_DispCapture_BlendToCustomDstBuffer(const u16 *srcA,
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < length; i++)
 	{
 		const u16 colorA = (!CAPTUREFROMNATIVESRCA) ? srcA[i] : srcA[offset + i];
@ -7095,6 +7126,9 @@ void ConvertColorBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restric
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		dst[i] = ConvertColor555To8888Opaque<SWAP_RB>(src[i]);
@ -7127,6 +7161,9 @@ void ConvertColorBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restric
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		dst[i] = ConvertColor555To6665Opaque<SWAP_RB>(src[i]);
@ -7146,6 +7183,9 @@ void ConvertColorBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount)
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		dst[i] = ConvertColor8888To6665<SWAP_RB>(src[i]);
@ -7165,6 +7205,9 @@ void ConvertColorBuffer6665To8888(const u32 *src, u32 *dst, size_t pixCount)
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		dst[i] = ConvertColor6665To8888<SWAP_RB>(src[i]);
@ -7191,6 +7234,9 @@ void ConvertColorBuffer8888To5551(const u32 *__restrict src, u16 *__restrict dst
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		dst[i] = ConvertColor8888To5551<SWAP_RB>(src[i]);
@ -7217,6 +7263,9 @@ void ConvertColorBuffer6665To5551(const u32 *__restrict src, u16 *__restrict dst
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		dst[i] = ConvertColor6665To5551<SWAP_RB>(src[i]);
--- a/desmume/src/GPU.h
+++ b/desmume/src/GPU.h
@ -1802,7 +1802,7 @@ template <bool SWAP_RB>
 FORCEINLINE void ConvertColor555To6665Opaque(const __m128i src, __m128i &dstLo, __m128i &dstHi)
 {
 	// Conversion algorithm:
-	//    RGB   5-bit to 6-bit formula: dstRGB8 = (srcRGB5 << 1) | ((srcRGB5 >> 4) & 0x01)
+	//    RGB   5-bit to 6-bit formula: dstRGB6 = (srcRGB5 << 1) | ((srcRGB5 >> 4) & 0x01)
 	if (SWAP_RB)
 	{
 		dstLo =                     _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src, 17), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(src, 12), _mm_set1_epi32(0x00010000)));
@ -1836,6 +1836,9 @@ FORCEINLINE void ConvertColor555To6665Opaque(const __m128i src, __m128i &dstLo,
 template <bool SWAP_RB>
 FORCEINLINE __m128i ConvertColor8888To6665(const __m128i src)
 {
+	// Conversion algorithm:
+	//    RGB   8-bit to 6-bit formula: dstRGB6 = (srcRGB8 >> 2)
+	//    Alpha 8-bit to 6-bit formula: dstA5   = (srcA8   >> 3)
 	__m128i rgb;
 	const __m128i a = _mm_and_si128( _mm_srli_epi32(src, 3), _mm_set1_epi32(0x1F000000) );
 	
--- a/desmume/src/OGLRender.cpp
+++ b/desmume/src/OGLRender.cpp
@ -994,6 +994,10 @@ Render3DError OpenGLRenderer::_FlushFramebufferConvertOnCPU(const FragmentColor
 					_mm_store_si128( (__m128i *)(dstFramebuffer + i + 4), ConvertColor8888To6665<true>(srcColorHi) );
 					_mm_store_si128( (__m128i *)(dstRGBA5551 + i), ConvertColor8888To5551<true>(srcColorLo, srcColorHi) );
 				}
+#endif
+				
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 				for (; i < pixCount; i++)
 				{
@ -1025,6 +1029,10 @@ Render3DError OpenGLRenderer::_FlushFramebufferConvertOnCPU(const FragmentColor
 					_mm_store_si128( (__m128i *)(dstFramebuffer + i + 4), srcColorHi );
 					_mm_store_si128( (__m128i *)(dstRGBA5551 + i), ConvertColor8888To5551<true>(srcColorLo, srcColorHi) );
 				}
+#endif
+				
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 				for (; i < pixCount; i++)
 				{
@ -1064,6 +1072,10 @@ Render3DError OpenGLRenderer::_FlushFramebufferConvertOnCPU(const FragmentColor
 						_mm_store_si128( (__m128i *)(dstFramebuffer + iw + 4), ConvertColor8888To6665<true>(srcColorHi) );
 						_mm_store_si128( (__m128i *)(dstRGBA5551 + iw), ConvertColor8888To5551<true>(srcColorLo, srcColorHi) );
 					}
+#endif
+					
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 					for (; x < pixCount; x++, ir++, iw++)
 					{
@ -1105,6 +1117,10 @@ Render3DError OpenGLRenderer::_FlushFramebufferConvertOnCPU(const FragmentColor
 						_mm_store_si128( (__m128i *)(dstFramebuffer + iw + 4), srcColorHi );
 						_mm_store_si128( (__m128i *)(dstRGBA5551 + iw), ConvertColor8888To5551<true>(srcColorLo, srcColorHi) );
 					}
+#endif
+					
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
 #endif
 					for (; x < pixCount; x++, ir++, iw++)
 					{
--- a/desmume/src/rasterize.cpp
+++ b/desmume/src/rasterize.cpp
@ -2059,6 +2059,9 @@ Render3DError SoftRasterizerRenderer_SSE2::ClearUsingValues(const FragmentColor
 		_mm_stream_si128((__m128i *)(this->_framebufferAttributes->isTranslucentPoly + i), attrIsTranslucentPoly_vec128);
 	}
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < pixCount; i++)
 	{
 		this->_framebufferColor[i] = clearColor6665;
--- a/desmume/src/render3D.cpp
+++ b/desmume/src/render3D.cpp
@ -237,6 +237,9 @@ void FragmentAttributesBuffer::SetAll(const FragmentAttributes &attr)
 	}
 #endif
 	
+#ifdef ENABLE_SSE2
+#pragma LOOPVECTORIZE_DISABLE
+#endif
 	for (; i < count; i++)
 	{
 		this->SetAtIndex(i, attr);
--- a/desmume/src/types.h
+++ b/desmume/src/types.h
@ -35,13 +35,13 @@

 //enforce a constraint: gdb stub requires developer
 #if defined(GDB_STUB) && !defined(DEVELOPER)
-#define DEVELOPER
+	#define DEVELOPER
 #endif

 #ifdef DEVELOPER
-#define IF_DEVELOPER(X) X
+	#define IF_DEVELOPER(X) X
 #else
-#define IF_DEVELOPER(X)
+	#define IF_DEVELOPER(X)
 #endif

 #ifdef HOST_WINDOWS
@ -96,19 +96,19 @@
 //dont apply these to types without further testing. it only works portably here on declarations of variables
 //cant we find a pattern other people use more successfully?
 #if _MSC_VER >= 1900
-#define DS_ALIGN(X) alignas(X)
+	#define DS_ALIGN(X) alignas(X)
 #elif defined(_MSC_VER) || defined(__INTEL_COMPILER)
-#define DS_ALIGN(X) __declspec(align(X))
+	#define DS_ALIGN(X) __declspec(align(X))
 #elif defined(__GNUC__)
-#define DS_ALIGN(X) __attribute__ ((aligned (X)))
+	#define DS_ALIGN(X) __attribute__ ((aligned (X)))
 #else
-#define DS_ALIGN(X)
+	#define DS_ALIGN(X)
 #endif

 #ifdef HOST_64
-#define CACHE_ALIGN_SIZE 64
+	#define CACHE_ALIGN_SIZE 64
 #else
-#define CACHE_ALIGN_SIZE 32
+	#define CACHE_ALIGN_SIZE 32
 #endif

 //use this for example when you want a byte value to be better-aligned
@ -117,49 +117,59 @@
 //---------------------------------------------

 #ifdef __MINGW32__
-#define FASTCALL __attribute__((fastcall))
-#define ASMJIT_CALL_CONV kX86FuncConvGccFastCall
+	#define FASTCALL __attribute__((fastcall))
+	#define ASMJIT_CALL_CONV kX86FuncConvGccFastCall
 #elif defined (__i386__) && !defined(__clang__)
-#define FASTCALL __attribute__((regparm(3)))
-#define ASMJIT_CALL_CONV kX86FuncConvGccRegParm3
+	#define FASTCALL __attribute__((regparm(3)))
+	#define ASMJIT_CALL_CONV kX86FuncConvGccRegParm3
 #elif defined(_MSC_VER) || defined(__INTEL_COMPILER)
-#define FASTCALL
-#define ASMJIT_CALL_CONV kX86FuncConvDefault
+	#define FASTCALL
+	#define ASMJIT_CALL_CONV kX86FuncConvDefault
 #else
-#define FASTCALL
-#define ASMJIT_CALL_CONV kX86FuncConvDefault
+	#define FASTCALL
+	#define ASMJIT_CALL_CONV kX86FuncConvDefault
 #endif

 #ifdef _MSC_VER
-#define _CDECL_ __cdecl
+	#define _CDECL_ __cdecl
 #else
-#define _CDECL_
+	#define _CDECL_
 #endif

 #ifndef INLINE
-#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
-#define INLINE _inline
-#else
-#define INLINE inline
-#endif
+	#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
+		#define INLINE _inline
+	#else
+		#define INLINE inline
+	#endif
 #endif

 #ifndef FORCEINLINE
-#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
-#define FORCEINLINE __forceinline
-#define MSC_FORCEINLINE __forceinline
-#else
-#define FORCEINLINE inline __attribute__((always_inline)) 
-#define MSC_FORCEINLINE
-#endif
+	#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
+		#define FORCEINLINE __forceinline
+		#define MSC_FORCEINLINE __forceinline
+	#else
+		#define FORCEINLINE inline __attribute__((always_inline))
+		#define MSC_FORCEINLINE
+	#endif
 #endif

 #ifndef NOINLINE
-#ifdef __GNUC__
-#define NOINLINE __attribute__((noinline))
-#else
-#define NOINLINE
+	#ifdef __GNUC__
+		#define NOINLINE __attribute__((noinline))
+	#else
+		#define NOINLINE
+	#endif
 #endif
+
+#ifndef LOOPVECTORIZE_DISABLE
+	#if defined(_MSC_VER)
+		#define LOOPVECTORIZE_DISABLE loop(no_vector)
+	#elif defined(__clang__)
+		#define LOOPVECTORIZE_DISABLE clang loop vectorize(disable)
+	#else
+		#define LOOPVECTORIZE_DISABLE
+	#endif
 #endif

 #if defined(__LP64__)