VideoCommon: merged SSSE3/SSE4.1 codes. Added some additional SSSE3/SSE4.1 codes which will be used in "The Legend of Zelda: Twilight Princess".

These codes don't work unless "_M_SSE=0x301", for SSSE3, or "_M_SSE=0x401", for SSE4.1, is defined as a preprocessor definition.


git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5300 8ced0084-cf51-0410-be5f-012b33b47a6e

4 files changed, 171 insertions, 5 deletions

diff --git a/Source/Core/VideoCommon/Src/TextureDecoder.cpp b/Source/Core/VideoCommon/Src/TextureDecoder.cpp
index ecd78ba16b..7cc0c05247 100644
--- a/Source/Core/VideoCommon/Src/TextureDecoder.cpp
+++ b/Source/Core/VideoCommon/Src/TextureDecoder.cpp
@@ -16,6 +16,13 @@
 // http://code.google.com/p/dolphin-emu/
 
 #include <cmath>
+
+#if _M_SSE >= 0x401
+#include <smmintrin.h>
+#elif _M_SSE >= 0x301
+#include <tmmintrin.h>
+#endif
+
 #include "Common.h"
 //#include "VideoCommon.h" // to get debug logs
 
@@ -359,14 +366,35 @@ inline void decodebytesC8_5A3_To_BGRA32(u32 *dst, const u8 *src, int tlutaddr)
     }
 }
 
+#if _M_SSE >= 0x301
+	static const __m128i kMaskSwap16 = _mm_set_epi32(0x0E0F0C0DL, 0x0A0B0809L, 0x06070405L, 0x02030001L);
+#endif
+
 inline void decodebytesC8_To_Raw16(u16* dst, const u8* src, int tlutaddr)
 {
 	u16* tlut = (u16*)(texMem + tlutaddr);
+
+#if _M_SSE >= 0x301
+
+	// Make 8 16-bits unsigned integer values
+	const __m128i a = _mm_set_epi16(tlut[src[7]], tlut[src[6]], tlut[src[5]], tlut[src[4]], tlut[src[3]], tlut[src[2]], tlut[src[1]], tlut[src[0]]);
+
+	// Apply Common::swap16() to 16-bits unsigned integers at once
+	const __m128i b = _mm_shuffle_epi8(a, kMaskSwap16);
+
+	// Store values to dst without polluting the caches
+	_mm_stream_si128((__m128i*)dst, b);
+
+#else
+
 	for (int x = 0; x < 8; x++)
 	{
 		u8 val = src[x];
 		*dst++ = Common::swap16(tlut[val]);
 	}
+
+#endif
+
 }
 
 
@@ -869,7 +897,9 @@ PC_TexFormat TexDecoder_DirectDecode_real(u8 *dst, const u8 *src, int width, int
     return PC_TEX_FMT_NONE;
 }
 
-
+#if _M_SSE >= 0x301
+static const __m128i kMaskSwap32 = _mm_set_epi32(0x0C0D0E0FL, 0x08090A0BL, 0x04050607L, 0x00010203L);
+#endif
 
 //switch endianness, unswizzle
 //TODO: to save memory, don't blindly convert everything to argb8888
@@ -997,13 +1027,59 @@ PC_TexFormat TexDecoder_Decode_real(u8 *dst, const u8 *src, int width, int heigh
         return PC_TEX_FMT_BGRA32;
     case GX_TF_RGBA8:  // speed critical
         {
+
+#if _M_SSE >= 0x301
+
+			for (int y = 0; y < height; y += 4) {
+				__m128i* p = (__m128i*)(src + y * width * 4);
+				for (int x = 0; x < width; x += 4) {
+
+#if _M_SSE >= 0x401
+					// Load 64-bytes at once.
+					const __m128i a0 = _mm_stream_load_si128(p++);
+					const __m128i a1 = _mm_stream_load_si128(p++);
+					const __m128i a2 = _mm_stream_load_si128(p++);
+					const __m128i a3 = _mm_stream_load_si128(p++);
+#else
+					const __m128i a0 = _mm_load_si128(p++);
+					const __m128i a1 = _mm_load_si128(p++);
+					const __m128i a2 = _mm_load_si128(p++);
+					const __m128i a3 = _mm_load_si128(p++);
+#endif
+
+					// Shuffle 16-bit integeres by _mm_unpacklo_epi16()/_mm_unpackhi_epi16(),
+					// apply Common::swap32() by _mm_shuffle_epi8() and
+					// store them by _mm_stream_si128().
+					// See decodebytesARGB8_4() about the idea.
+					const __m128i b0 = _mm_unpacklo_epi16(a0, a2);
+					const __m128i c0 = _mm_shuffle_epi8(b0, kMaskSwap32);
+					_mm_stream_si128((__m128i*)((u32*)dst + (y + 0) * width + x), c0);
+
+					const __m128i b1 = _mm_unpackhi_epi16(a0, a2);
+					const __m128i c1 = _mm_shuffle_epi8(b1, kMaskSwap32);
+					_mm_stream_si128((__m128i*)((u32*)dst + (y + 1) * width + x), c1);
+
+					const __m128i b2 = _mm_unpacklo_epi16(a1, a3);
+					const __m128i c2 = _mm_shuffle_epi8(b2, kMaskSwap32);
+					_mm_stream_si128((__m128i*)((u32*)dst + (y + 2) * width + x), c2);
+
+					const __m128i b3 = _mm_unpackhi_epi16(a1, a3);
+					const __m128i c3 = _mm_shuffle_epi8(b3, kMaskSwap32);
+					_mm_stream_si128((__m128i*)((u32*)dst + (y + 3) * width + x), c3);
+				}
+			}
+
+#else
+
 			for (int y = 0; y < height; y += 4)
-                for (int x = 0; x < width; x += 4)
-                {
+				for (int x = 0; x < width; x += 4)
+				{
 					for (int iy = 0; iy < 4; iy++)
-                        decodebytesARGB8_4((u32*)dst + (y+iy)*width + x, (u16*)src + 4 * iy, (u16*)src + 4 * iy + 16);
+						decodebytesARGB8_4((u32*)dst + (y+iy)*width + x, (u16*)src + 4 * iy, (u16*)src + 4 * iy + 16);
 					src += 64;
-                }
+				}
+
+#endif
         }
         return PC_TEX_FMT_BGRA32;
     case GX_TF_CMPR:  // speed critical
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
index bf38ef1e80..4cee83d9f6 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
@@ -21,6 +21,10 @@
 #include "VertexLoader_Normal.h"
 #include "NativeVertexWriter.h"
 
+#if _M_SSE >= 0x301
+#include <tmmintrin.h>
+#endif
+
 #define LOG_NORM8() // PRIM_LOG("norm: %f %f %f, ", ((s8*)VertexManager::s_pCurBufferPointer)[-3]/127.0f, ((s8*)VertexManager::s_pCurBufferPointer)[-2]/127.0f, ((s8*)VertexManager::s_pCurBufferPointer)[-1]/127.0f);
 #define LOG_NORM16() // PRIM_LOG("norm: %f %f %f, ", ((s16*)VertexManager::s_pCurBufferPointer)[-3]/32767.0f, ((s16*)VertexManager::s_pCurBufferPointer)[-2]/32767.0f, ((s16*)VertexManager::s_pCurBufferPointer)[-1]/32767.0f);
 #define LOG_NORMF() // PRIM_LOG("norm: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
@@ -411,14 +415,30 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte_Expand16()
     LOG_NORM16();
 }
 
+#if _M_SSE >= 0x301
+static const __m128i kMaskSwap16_3 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x0FFFF0405L, 0x02030001L);
+#endif
+
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Short()
 {
     u16 Index = DataReadU16();
 	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
+
+#if _M_SSE >= 0x301
+
+    __m128i a = _mm_loadl_epi64((__m128i*)pData);
+    __m128i b = _mm_shuffle_epi8(a, kMaskSwap16_3);
+    _mm_storel_epi64((__m128i*)VertexManager::s_pCurBufferPointer, b);
+
+#else
+
     ((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(pData[0]);
     ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(pData[1]);
     ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(pData[2]);
     ((u16*)VertexManager::s_pCurBufferPointer)[3] = 0;
+
+#endif
+
     VertexManager::s_pCurBufferPointer += 8;
     LOG_NORM16();
 }
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
index aa193d4955..6ff0884cec 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
@@ -21,6 +21,10 @@
 #include "VertexLoader_Position.h"
 #include "NativeVertexWriter.h"
 
+#if _M_SSE >= 301
+#include <tmmintrin.h>
+#endif
+
 extern float posScale;
 extern TVtxAttr *pVtxAttr;
 
@@ -146,16 +150,33 @@ inline void Pos_ReadIndex_Short(int Index)
 	VertexManager::s_pCurBufferPointer += 12;
 }
 
+#if _M_SSE >= 0x301
+static const __m128i kMaskSwap32_3 = _mm_set_epi32(0xFFFFFFFFL, 0x08090A0BL, 0x04050607L, 0x00010203L);
+static const __m128i kMaskSwap32_2 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x04050607L, 0x00010203L);
+#endif
+
 template<bool three>
 inline void Pos_ReadIndex_Float(int Index)
 {
 	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_POSITION] + (Index * arraystrides[ARRAY_POSITION]));
+
+#if _M_SSE >= 0x301
+
+	const __m128i a = _mm_loadu_si128((__m128i*)pData);
+	__m128i b = _mm_shuffle_epi8(a, three ? kMaskSwap32_3 : kMaskSwap32_2);
+	_mm_storeu_si128((__m128i*)VertexManager::s_pCurBufferPointer, b);
+
+#else
+
 	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(pData[0]);
 	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(pData[1]);
 	if (three)
 		((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(pData[2]);
 	else
 		((float*)VertexManager::s_pCurBufferPointer)[2] = 0.0f;
+
+#endif
+
 	LOG_VTX();
 	VertexManager::s_pCurBufferPointer += 12;
 }
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp b/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
index e6720e81a0..ae4e6d0bb8 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
@@ -21,6 +21,12 @@
 #include "VertexLoader_TextCoord.h"
 #include "NativeVertexWriter.h"
 
+#if _M_SSE >= 0x401
+#include <smmintrin.h>
+#elif _M_SSE >= 0x301
+#include <tmmintrin.h>
+#endif
+
 #define LOG_TEX1() // PRIM_LOG("tex: %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0]);
 #define LOG_TEX2() // PRIM_LOG("tex: %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0], ((float*)VertexManager::s_pCurBufferPointer)[1]);
 
@@ -284,13 +290,35 @@ void LOADERDECL TexCoord_ReadIndex16_Short1()
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
 }
+
+#if _M_SSE >= 0x401
+static const __m128i kMaskSwap16_2 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x02030001L);
+#endif
+
 void LOADERDECL TexCoord_ReadIndex16_Short2()	
 {
 	// Heavy in ZWW
 	u16 Index = DataReadU16(); 
+
+#if _M_SSE >= 0x401
+
+	const s32 *pData = (const s32*)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	const __m128i a = _mm_cvtsi32_si128(*pData);
+	const __m128i b = _mm_shuffle_epi8(a, kMaskSwap16_2);
+	const __m128i c = _mm_cvtepi16_epi32(b);
+	const __m128 d = _mm_cvtepi32_ps(c);
+	const __m128 e = _mm_load1_ps(&tcScale[tcIndex]);
+	const __m128 f = _mm_mul_ps(d, e);
+	_mm_storeu_ps((float*)VertexManager::s_pCurBufferPointer, f);
+
+#else
+
 	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
 	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Common::swap16(pData[0]) * tcScale[tcIndex];
 	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s16)Common::swap16(pData[1]) * tcScale[tcIndex];
+
+#endif
+
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -305,15 +333,36 @@ void LOADERDECL TexCoord_ReadIndex16_Float1()
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
 }
+
+#if _M_SSE >= 0x301
+static const __m128i kMaskSwap32 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x04050607L, 0x00010203L);
+#endif
+
 void LOADERDECL TexCoord_ReadIndex16_Float2()	
 {
 	u16 Index = DataReadU16(); 
 	const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+
+#if _M_SSE >= 0x301
+
+	const __m128i a = _mm_loadl_epi64((__m128i*)pData);
+	const __m128i b = _mm_shuffle_epi8(a, kMaskSwap32);
+	u8* p = VertexManager::s_pCurBufferPointer;
+	_mm_storel_epi64((__m128i*)p, b);
+	LOG_TEX2();
+	p += 8;
+	VertexManager::s_pCurBufferPointer = p;
+	tcIndex++;
+
+#else
+
 	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(pData[0]);
 	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(pData[1]);
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
+
+#endif
 }
 
 ReadTexCoord tableReadTexCoord[4][8][2] = {