TextureDecoder.cpp: new SSE2 optimized GX_TF_I8 decoder. Probably not ultimately optimal SSE2 code, but provably better (on my machine) than the memset version. Tested with __rdtsc counts in an independent project. I get about 6-7 FPS more on average during the intro movie playback in Mario Kart Wii. Hope this compiles for GCC okay.

TextureDecoder.cpp: merged two functionally identical decode5A3RGBA and decode5A3rgba methods.
OpcodeDecoding.cpp and DLCache.cpp: optimization for GX_LOAD_XF_REG. The PSUHFB solution sounds better for SSSE3, but this is a small win for the default case.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6692 8ced0084-cf51-0410-be5f-012b33b47a6e

5 files changed, 158 insertions, 60 deletions

diff --git a/Source/Core/VideoCommon/Src/DLCache.cpp b/Source/Core/VideoCommon/Src/DLCache.cpp
index 932d561580..4d76cfb290 100644
--- a/Source/Core/VideoCommon/Src/DLCache.cpp
+++ b/Source/Core/VideoCommon/Src/DLCache.cpp
@@ -307,8 +307,7 @@ u8 AnalyzeAndRunDisplayList(u32 address, int	 size, CachedDisplayList *dl)
 					u32 xf_address = Cmd2 & 0xFFFF;
 					// TODO - speed this up. pshufb?
 					u32 data_buffer[16];
-					for (int i = 0; i < transfer_size; i++)
-						data_buffer[i] = DataReadU32();
+					DataReadU32xFuncs[transfer_size-1](data_buffer);
 					LoadXFReg(transfer_size, xf_address, data_buffer);
 					INCSTAT(stats.thisFrame.numXFLoads);
 					num_xf_reg++;
@@ -462,8 +461,7 @@ bool CompileAndRunDisplayList(u32 address, int size, CachedDisplayList *dl)
 					NewRegion->hash = 0;					
 					dl->InsertRegion(NewRegion);
 					u32 *data_buffer = (u32*)NewRegion->start_address;
-					for (int i = 0; i < transfer_size; i++)
-						data_buffer[i] = DataReadU32();
+					DataReadU32xFuncs[transfer_size-1](data_buffer);
 					LoadXFReg(transfer_size, xf_address, data_buffer);
 					INCSTAT(stats.thisFrame.numXFLoads);
 					// Compile
diff --git a/Source/Core/VideoCommon/Src/DataReader.h b/Source/Core/VideoCommon/Src/DataReader.h
index f98736c407..933de18849 100644
--- a/Source/Core/VideoCommon/Src/DataReader.h
+++ b/Source/Core/VideoCommon/Src/DataReader.h
@@ -64,6 +64,32 @@ __forceinline u32 DataReadU32()
 	return tmp;
 }
 
+template<unsigned int N>
+void DataReadU32xN(u32 *bufx16)
+{
+	memcpy(bufx16, g_pVideoData, sizeof(u32) * N);
+	if (N >= 1) bufx16[0] = Common::swap32(bufx16[0]);
+	if (N >= 2) bufx16[1] = Common::swap32(bufx16[1]);
+	if (N >= 3) bufx16[2] = Common::swap32(bufx16[2]);
+	if (N >= 4) bufx16[3] = Common::swap32(bufx16[3]);
+	if (N >= 5) bufx16[4] = Common::swap32(bufx16[4]);
+	if (N >= 6) bufx16[5] = Common::swap32(bufx16[5]);
+	if (N >= 7) bufx16[6] = Common::swap32(bufx16[6]);
+	if (N >= 8) bufx16[7] = Common::swap32(bufx16[7]);
+	if (N >= 9) bufx16[8] = Common::swap32(bufx16[8]);
+	if (N >= 10) bufx16[9] = Common::swap32(bufx16[9]);
+	if (N >= 11) bufx16[10] = Common::swap32(bufx16[10]);
+	if (N >= 12) bufx16[11] = Common::swap32(bufx16[11]);
+	if (N >= 13) bufx16[12] = Common::swap32(bufx16[12]);
+	if (N >= 14) bufx16[13] = Common::swap32(bufx16[13]);
+	if (N >= 15) bufx16[14] = Common::swap32(bufx16[14]);
+	if (N >= 16) bufx16[15] = Common::swap32(bufx16[15]);
+	g_pVideoData += (sizeof(u32) * N);
+}
+
+typedef void (*DataReadU32xNfunc)(u32 *buf);
+extern DataReadU32xNfunc DataReadU32xFuncs[16];
+
 __forceinline u32 DataReadU32Unswapped()
 {
 	u32 tmp = *(u32*)g_pVideoData;
diff --git a/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp b/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp
index 4fb9574478..fbb81fb40d 100644
--- a/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp
+++ b/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp
@@ -49,6 +49,24 @@
 #endif
 
 u8* g_pVideoData = 0;
+DataReadU32xNfunc DataReadU32xFuncs[16] = {
+	DataReadU32xN<1>,
+	DataReadU32xN<2>,
+	DataReadU32xN<3>,
+	DataReadU32xN<4>,
+	DataReadU32xN<5>,
+	DataReadU32xN<6>,
+	DataReadU32xN<7>,
+	DataReadU32xN<8>,
+	DataReadU32xN<9>,
+	DataReadU32xN<10>,
+	DataReadU32xN<11>,
+	DataReadU32xN<12>,
+	DataReadU32xN<13>,
+	DataReadU32xN<14>,
+	DataReadU32xN<15>,
+	DataReadU32xN<16>
+};
 
 extern u8* FAKE_GetFifoStartPtr();
 extern u8* FAKE_GetFifoEndPtr();
@@ -233,12 +251,13 @@ static void Decode()
         {
             u32 Cmd2 = DataReadU32();
             int transfer_size = ((Cmd2 >> 16) & 15) + 1;
-            u32 xf_address = Cmd2 & 0xFFFF;
+			u32 xf_address = Cmd2 & 0xFFFF;
 			// TODO - speed this up. pshufb?
-            u32 data_buffer[16];
-            for (int i = 0; i < transfer_size; i++)
-                data_buffer[i] = DataReadU32();
+			u32 data_buffer[16];
+			DataReadU32xFuncs[transfer_size-1](data_buffer);
+
             LoadXFReg(transfer_size, xf_address, data_buffer);
+
 			INCSTAT(stats.thisFrame.numXFLoads);
         }
         break;
@@ -317,7 +336,7 @@ static void DecodeSemiNop()
             u8 sub_cmd = DataReadU8();
             u32 value = DataReadU32();
             LoadCPReg(sub_cmd, value);
-			INCSTAT(stats.thisFrame.numCPLoads);
+            INCSTAT(stats.thisFrame.numCPLoads);
         }
         break;
 
@@ -328,10 +347,9 @@ static void DecodeSemiNop()
             u32 address = Cmd2 & 0xFFFF;
 			// TODO - speed this up. pshufb?
             u32 data_buffer[16];
-            for (int i = 0; i < transfer_size; i++)
-                data_buffer[i] = DataReadU32();
+            DataReadU32xFuncs[transfer_size-1](data_buffer);
             LoadXFReg(transfer_size, address, data_buffer);
-			INCSTAT(stats.thisFrame.numXFLoads);
+            INCSTAT(stats.thisFrame.numXFLoads);
         }
         break;
 
diff --git a/Source/Core/VideoCommon/Src/TextureDecoder.cpp b/Source/Core/VideoCommon/Src/TextureDecoder.cpp
index dd350d45c7..af29d6b4b4 100644
--- a/Source/Core/VideoCommon/Src/TextureDecoder.cpp
+++ b/Source/Core/VideoCommon/Src/TextureDecoder.cpp
@@ -32,6 +32,7 @@
 
 #if _M_SSE >= 0x401
 #include <smmintrin.h>
+#include <emmintrin.h>
 #elif _M_SSE >= 0x301 && !(defined __GNUC__ && !defined __SSSE3__)
 #include <tmmintrin.h>
 #endif
@@ -209,24 +210,24 @@ inline u32 decode5A3(u16 val)
     return (a << 24) | (r << 16) | (g << 8) | b;
 }
 
-inline u32 decode5A3rgba(u16 val)
+inline u32 decode5A3RGBA(u16 val)
 {
     int r,g,b,a;
-    if ((val & 0x8000))
+    if ((val&0x8000))
     {
-        a = 0xFF;
-		r = Convert5To8((val >> 10) & 0x1F);
-		g = Convert5To8((val >> 5) & 0x1F);
-		b = Convert5To8(val & 0x1F);
+        r=Convert5To8((val>>10) & 0x1f);
+        g=Convert5To8((val>>5 ) & 0x1f);
+        b=Convert5To8((val    ) & 0x1f);
+        a=0xFF;
 	}
     else
     {
-		a = Convert3To8((val >> 12) & 0x7);
-		r = Convert4To8((val >> 8) & 0xF);
-		g = Convert4To8((val >> 4) & 0xF);
-		b = Convert4To8(val & 0xF);
+        a=Convert3To8((val>>12) & 0x7);
+        r=Convert4To8((val>>8 ) & 0xf);
+        g=Convert4To8((val>>4 ) & 0xf);
+        b=Convert4To8((val    ) & 0xf);
     }
-    return (a << 24) | (b << 16) | (g << 8) | r;
+    return r | (g<<8) | (b << 16) | (a << 24);
 }
 
 inline u32 decode565RGBA(u16 val)
@@ -246,26 +247,6 @@ inline u32 decodeIA8Swapped(u16 val)
     return i | (i<<8) | (i<<16) | (a<<24);
 }
 
-inline u32 decode5A3RGBA(u16 val)
-{
-    int r,g,b,a;
-    if ((val&0x8000))
-    {
-        r=Convert5To8((val>>10) & 0x1f);
-        g=Convert5To8((val>>5 ) & 0x1f);
-        b=Convert5To8((val    ) & 0x1f);
-        a=0xFF;
-	}
-    else
-    {
-        a=Convert3To8((val>>12) & 0x7);
-        r=Convert4To8((val>>8 ) & 0xf);
-        g=Convert4To8((val>>4 ) & 0xf);
-        b=Convert4To8((val    ) & 0xf);
-    }
-    return r | (g<<8) | (b << 16) | (a << 24);
-}
-
 
 
 struct DXTBlock
@@ -293,8 +274,8 @@ inline void decodebytesC4_5A3_To_rgba32(u32 *dst, const u8 *src, int tlutaddr)
     for (int x = 0; x < 4; x++)
     {
         u8 val = src[x];
-        *dst++ = decode5A3rgba(Common::swap16(tlut[val >> 4]));
-        *dst++ = decode5A3rgba(Common::swap16(tlut[val & 0xF]));
+        *dst++ = decode5A3RGBA(Common::swap16(tlut[val >> 4]));
+        *dst++ = decode5A3RGBA(Common::swap16(tlut[val & 0xF]));
     }
 }
 
@@ -348,7 +329,7 @@ inline void decodebytesC8_5A3_To_RGBA32(u32 *dst, const u8 *src, int tlutaddr)
     for (int x = 0; x < 8; x++)
     {
         u8 val = src[x];
-        *dst++ = decode5A3rgba(Common::swap16(tlut[val]));
+        *dst++ = decode5A3RGBA(Common::swap16(tlut[val]));
     }
 }
 
@@ -422,7 +403,7 @@ inline void decodebytesC14X2_5A3_To_RGBA(u32 *dst, const u16 *src, int tlutaddr)
     for (int x = 0; x < 4; x++)
     {
         u16 val = Common::swap16(src[x]);
-		*dst++ = decode5A3rgba(Common::swap16(tlut[(val & 0x3FFF)]));
+		*dst++ = decode5A3RGBA(Common::swap16(tlut[(val & 0x3FFF)]));
     }
 }
 
@@ -481,23 +462,43 @@ inline void decodebytesIA4RGBA(u32 *dst, const u8 *src)
 
 inline void decodebytesRGB5A3(u32 *dst, const u16 *src)
 {
+#if 0
     for (int x = 0; x < 4; x++)
         dst[x] = decode5A3(Common::swap16(src[x]));
+#else
+    dst[0] = decode5A3(Common::swap16(src[0]));
+    dst[1] = decode5A3(Common::swap16(src[1]));
+    dst[2] = decode5A3(Common::swap16(src[2]));
+    dst[3] = decode5A3(Common::swap16(src[3]));
+#endif
 }
 
 inline void decodebytesRGB5A3rgba(u32 *dst, const u16 *src)
 {
+#if 0
     for (int x = 0; x < 4; x++)
-        dst[x] = decode5A3rgba(Common::swap16(src[x]));
+        dst[x] = decode5A3RGBA(Common::swap16(src[x]));
+#else
+    dst[0] = decode5A3RGBA(Common::swap16(src[0]));
+    dst[1] = decode5A3RGBA(Common::swap16(src[1]));
+    dst[2] = decode5A3RGBA(Common::swap16(src[2]));
+    dst[3] = decode5A3RGBA(Common::swap16(src[3]));
+#endif
 }
 
 // This one is used by many video formats. It'd therefore be good if it was fast.
 // Needs more speed.
 inline void decodebytesARGB8_4(u32 *dst, const u16 *src, const u16 *src2)
 {
-    for (int x = 0; x < 4; x++) {
+#if 0
+    for (int x = 0; x < 4; x++)
         dst[x] = Common::swap32((src2[x] << 16) | src[x]);
-    }
+#else
+    dst[0] = Common::swap32((src2[0] << 16) | src[0]);
+    dst[1] = Common::swap32((src2[1] << 16) | src[1]);
+    dst[2] = Common::swap32((src2[2] << 16) | src[2]);
+    dst[3] = Common::swap32((src2[3] << 16) | src[3]);
+#endif
 
 	// This can probably be done in a few SSE pack/unpack instructions + pshufb
 	// some unpack instruction x2:
@@ -508,11 +509,18 @@ inline void decodebytesARGB8_4(u32 *dst, const u16 *src, const u16 *src2)
 	// and we are done.
 }
 
-inline void decodebytesARGB8_4ToRgba(u32 *dst, const u16 *src, const u16 *src2)
+inline void decodebytesARGB8_4ToRgba(u32 *dst, const u16 *src, const u16 * src2)
 {
-	for (int x = 0; x < 4; x++) {
+#if 0
+    for (int x = 0; x < 4; x++) {
 		dst[x] =  ((src[x] & 0xFF) << 24) | ((src[x] & 0xFF00)>>8)  | (src2[x] << 8);		
-    }	 
+	}
+#else
+    dst[0] =  ((src[0] & 0xFF) << 24) | ((src[0] & 0xFF00)>>8)  | (src2[0] << 8);
+    dst[1] =  ((src[1] & 0xFF) << 24) | ((src[1] & 0xFF00)>>8)  | (src2[1] << 8);
+    dst[2] =  ((src[2] & 0xFF) << 24) | ((src[2] & 0xFF00)>>8)  | (src2[2] << 8);
+    dst[3] =  ((src[3] & 0xFF) << 24) | ((src[3] & 0xFF00)>>8)  | (src2[3] << 8);
+#endif
 }
 
 inline u32 makecol(int r, int g, int b, int a)
@@ -919,7 +927,7 @@ PC_TexFormat TexDecoder_Decode_real(u8 *dst, const u8 *src, int width, int heigh
 
 
 
-PC_TexFormat TexDecoder_Decode_RGBA(u32 *dst, const u8 *src, int width, int height, int texformat, int tlutaddr, int tlutfmt)
+PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int height, int texformat, int tlutaddr, int tlutfmt)
 {
     switch (texformat)
     {
@@ -966,9 +974,52 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 *dst, const u8 *src, int width, int heig
 		{
 			for (int y = 0; y < height; y += 4)
 				for (int x = 0; x < width; x += 8)
-					for (int iy = 0; iy < 4; iy++, src += 8)
-						for (int ix = 0; ix < 8; ix++)
-							memset(dst + (y + iy)*width+x+ ix, (src + ix)[0], 4);
+#if _M_SSE >= 0x401
+
+                    for (int iy = 0; iy < 4; ++iy, src += 8)
+                    {
+                        __m128i *quaddst = (__m128i *)(dst + (y + iy)*width + x);
+                        const __m128i m0 = _mm_or_si128(
+                            _mm_or_si128(
+                                _mm_and_si128(_mm_set1_epi8(src[0]), _mm_set_epi32(0, 0, 0, (int)0xffffffffU)),
+                                _mm_and_si128(_mm_set1_epi8(src[1]), _mm_set_epi32(0, 0, (int)0xffffffffU, 0))
+                            ),
+                            _mm_or_si128(
+                                _mm_and_si128(_mm_set1_epi8(src[2]), _mm_set_epi32(0, (int)0xffffffffU, 0, 0)),
+                                _mm_and_si128(_mm_set1_epi8(src[3]), _mm_set_epi32((int)0xffffffffU, 0, 0, 0))
+                            )
+                        );
+                        _mm_store_si128(quaddst, m0);
+
+                        const __m128i m1 = _mm_or_si128(
+                            _mm_or_si128(
+                                _mm_and_si128(_mm_set1_epi8(src[4]), _mm_set_epi32(0, 0, 0, (int)0xffffffffU)),
+                                _mm_and_si128(_mm_set1_epi8(src[5]), _mm_set_epi32(0, 0, (int)0xffffffffU, 0))
+                            ),
+                            _mm_or_si128(
+                                _mm_and_si128(_mm_set1_epi8(src[6]), _mm_set_epi32(0, (int)0xffffffffU, 0, 0)),
+                                _mm_and_si128(_mm_set1_epi8(src[7]), _mm_set_epi32((int)0xffffffffU, 0, 0, 0))
+                            )
+                        );
+                        _mm_store_si128(quaddst+1, m1);
+                    }
+#else
+                    for (int iy = 0; iy < 4; ++iy, src += 8)
+                    {
+                        u32 *  newdst = dst + (y + iy)*width+x;
+                        const u8 *  newsrc = src;
+                        u8 srcval;
+
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = (newsrc++)[0]; (newdst++)[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                        srcval = newsrc[0]; newdst[0] = srcval | (srcval << 8) | (srcval << 16) | (srcval << 24);
+                    }
+#endif
 		}
 		break;
     case GX_TF_C8:
@@ -1014,8 +1065,10 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 *dst, const u8 *src, int width, int heig
 					{
 						u32 *ptr = dst + (y + iy) * width + x;
 						u16 *s = (u16 *)src;
-						for(int j = 0; j < 4; j++)
-							*ptr++ = decodeIA8Swapped(*s++);
+						ptr[0] = decodeIA8Swapped(s[0]);
+						ptr[1] = decodeIA8Swapped(s[1]);
+						ptr[2] = decodeIA8Swapped(s[2]);
+						ptr[3] = decodeIA8Swapped(s[3]);
 					}
 
         }
@@ -1058,7 +1111,7 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 *dst, const u8 *src, int width, int heig
 		}
 		break;
     case GX_TF_RGB5A3:
-        {
+        {   // JSD: speed critical for Mario Kart Wii intro movie (at least)
             for (int y = 0; y < height; y += 4)
                 for (int x = 0; x < width; x += 4)
                     for (int iy = 0; iy < 4; iy++, src += 8)
diff --git a/Source/Core/VideoCommon/VideoCommon.vcproj b/Source/Core/VideoCommon/VideoCommon.vcproj
index 7685d1bba8..e30905f34a 100644
--- a/Source/Core/VideoCommon/VideoCommon.vcproj
+++ b/Source/Core/VideoCommon/VideoCommon.vcproj
@@ -387,6 +387,9 @@
 			/>
 			<Tool
 				Name="VCCLCompilerTool"
+				Optimization="2"
+				InlineFunctionExpansion="1"
+				EnableIntrinsicFunctions="true"
 				AdditionalIncludeDirectories="../../PluginSpecs;../../../Externals/CLRun/include;../../../Externals/SOIL;../Common/Src;../Core/Src"
 				PreprocessorDefinitions="WIN32;NDEBUG;_LIB;_SECURE_SCL=0;__WXMSW__;wxUSE_BASE=0;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE"
 				RuntimeLibrary="0"