Substantial XFB speedup. There's more to get though, for example by using a shader for color conversion instead - but i like having a fast CPU implementation too. Also adds some sanity checks. PAL games still have problems.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@885 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Core/Src/HW/VideoInterface.cpp

@@ -508,25 +508,32 @@ void Update()
             VerticalBeamPos = 1;
         }
 
-		if(VerticalBeamPos == NextXFBRender)
+		if (VerticalBeamPos == NextXFBRender)
 		{
 			
 			u8* xfbPtr = 0;
 			int yOffset = 0;
 
-			if(NextXFBRender == 1)
+			if (NextXFBRender == 1)
 			{
 				NextXFBRender = LinesPerField;
-				xfbPtr = Memory::GetPointer(VideoInterface::m_FrameBufferTop.Hex);
+				// The & mask is a hack for mario kart
+				u32 addr = (VideoInterface::m_FrameBufferTop.Hex & 0xFFFFFFF) | 0x80000000;
+				if (addr >= 0x80000000 && 
+					addr <= (0x81800000-640*480*2))
+					xfbPtr = Memory::GetPointer(addr);
 			}
 			else
 			{
 				NextXFBRender = 1;
-				xfbPtr = Memory::GetPointer(VideoInterface::m_FrameBufferBottom.Hex);
+				u32 addr = (VideoInterface::m_FrameBufferBottom.Hex & 0xFFFFFFF) | 0x80000000;
+				if (addr >= 0x80000000 && 
+					addr <= (0x81800000-640*480*2))
+					xfbPtr = Memory::GetPointer(addr);
 				yOffset = -1;
 			}
 
-			if(xfbPtr && PluginVideo::IsLoaded())
+			if (xfbPtr && PluginVideo::IsLoaded())
 			{
 				int fbWidth = m_VIHorizontalStepping.FieldSteps * 16;
 				int fbHeight = (m_VIHorizontalStepping.FbSteps / m_VIHorizontalStepping.FieldSteps) * m_VIVerticalTimingRegister.ACV;				
@@ -549,6 +556,6 @@ void Update()
         }
     }
 }
-
+ 
 }
 

Source/Core/DolphinWX/Src/Frame.cpp

@@ -347,7 +347,7 @@ void CFrame::OnOpen(wxCommandEvent& WXUNUSED (event))
 			wxEmptyString, wxEmptyString, wxEmptyString,
 			wxString::Format
 			(
-					_T("All GC/Wii files (elf, dol, gcm, iso)|*.elf;*.dol;*.gcm;*.iso|All files (%s)|%s"),
+					_T("All GC/Wii files (elf, dol, gcm, iso)|*.elf;*.dol;*.gcm;*.iso;*.gcz|All files (%s)|%s"),
 					wxFileSelectorDefaultWildcardStr,
 					wxFileSelectorDefaultWildcardStr
 			),
@@ -607,6 +607,9 @@ void CFrame::OnToggleStatusbar(wxCommandEvent& event)
 
 void CFrame::OnKeyDown(wxKeyEvent& event)
 {
+	event.Skip();
+	return;
+
 	if (((event.GetKeyCode() == WXK_RETURN) && (event.GetModifiers() == wxMOD_ALT)) ||
 	    (event.GetKeyCode() == WXK_ESCAPE))
 	{

Source/Core/VideoCommon/Src/XFBConvert.cpp

@@ -15,26 +15,39 @@
 // Official SVN repository and contact information can be found at
 // http://code.google.com/p/dolphin-emu/
 
+#if _WIN32
+#include <intrin.h>
+#endif
+
+#include <xmmintrin.h>
+
 #include "XFBConvert.h"
 #include "Common.h"
 
-// TODO: Convert this thing into wicked fast SSE2.
-
 namespace {
 
-int bound(int i)
+u8 bound_table[3*256];
+int y[256];
+int v1[256];
+int v2[256];
+int u1[256];
+int u2[256];
+u8 *bound_lut = bound_table + 256;
+
+inline int bound(int i)
 {
-	return (i>255)?255:((i<0)?0:i);
+	return bound_lut[i];
 }
 
-void yuv2rgb(int y, int u, int v, int &r, int &g, int &b)
+inline void yuv2rgb(int y, int u, int v, int &r, int &g, int &b)
 {
-	b = bound((76283*(y - 16) + 132252*(u - 128))>>16);
+	int gray = 76283*(y - 16);
-	g = bound((76283*(y - 16) - 53281 *(v - 128) - 25624*(u - 128))>>16); //last one u?
+	b = bound((gray + 132252*(u - 128))>>16);
-	r = bound((76283*(y - 16) + 104595*(v - 128))>>16);
+	g = bound((gray - 53281 *(v - 128) - 25624*(u - 128))>>16);
+	r = bound((gray + 104595*(v - 128))>>16);
 }
 
-void rgb2yuv(int r, int g, int b, int &y, int &u, int &v)
+inline void rgb2yuv(int r, int g, int b, int &y, int &u, int &v)
 {
 	y = (((16843 * r) + (33030 * g) + (6423 * b)) >> 16) + 16;
 	v = (((28770 * r) - (24117 * g) - (4653 * b)) >> 16) + 128;
@@ -43,24 +56,61 @@ void rgb2yuv(int r, int g, int b, int &y, int &u, int &v)
 
 }  // namespace
 
+
+//const __m128i _bias1 = _mm_set_epi32(16 << 16, 128/2 << 16,        0, 128/2 << 16);
+//const __m128i _bias2 = _mm_set_epi32(0,        128/2 << 16, 16 << 16, 128/2 << 16);
+
+const __m128i _bias1 = _mm_set_epi32(128/2 << 16,        0, 128/2 << 16, 16 << 16);
+const __m128i _bias2 = _mm_set_epi32(128/2 << 16, 16 << 16, 128/2 << 16,        0);
+__m128i _r1[256];
+__m128i _r2[256];
+__m128i _g1[256];
+__m128i _g2[256];
+__m128i _b1[256];
+__m128i _b2[256];
+
+void InitXFBConvTables() 
+{
+	for (int i = 0; i < 256; i++)
+	{
+		bound_table[i] = 0;
+		bound_table[256 + i] = i;
+		bound_table[512 + i] = 255;
+
+		y[i] = 76283*(i - 16);
+		u1[i] = 132252 * (i - 128);
+		u2[i] = -25624 * (i - 128);
+		v1[i] = -53281 * (i - 128);
+		v2[i] = 104595 * (i - 128);
+
+		_r1[i] = _mm_set_epi32( 28770 * i / 2,          0,  -9699 * i / 2, 16843 * i);
+		_g1[i] = _mm_set_epi32(-24117 * i / 2,          0, -19071 * i / 2, 33030 * i);
+		_b1[i] = _mm_set_epi32( -4653 * i / 2,          0,  28770 * i / 2,  6423 * i);
+										   									
+		_r2[i] = _mm_set_epi32( 28770 * i / 2,  16843 * i,  -9699 * i / 2,         0);
+		_g2[i] = _mm_set_epi32(-24117 * i / 2,  33030 * i, -19071 * i / 2,         0);
+		_b2[i] = _mm_set_epi32( -4653 * i / 2,   6423 * i,  28770 * i / 2,         0);
+	}
+}
+
 void ConvertFromXFB(u32 *dst, const u8* _pXFB, int width, int height)
 {
 	const unsigned char *src = _pXFB;
-
 	u32 numBlocks = (width * height) / 2;
-
 	for (u32 i = 0; i < numBlocks; i++, src += 4)
 	{
-		int Y1 = src[0];
+		int Y1 = y[src[0]];
+		int Y2 = y[src[2]];
 		int U  = src[1];
-		int Y2 = src[2];
 		int V  = src[3];
-
+		int b1 = bound((Y1 + u1[U]) >> 16);
-		int r, g, b;
+		int g1 = bound((Y1 + v1[V] + u2[U])>>16);
-		yuv2rgb(Y1,U,V, r,g,b);
+		int r1 = bound((Y1 + v2[V]) >> 16);
-		*dst++ = 0xFF000000 | (r<<16) | (g<<8) | (b);
+		int b2 = bound((Y2 + u1[U]) >> 16);
-		yuv2rgb(Y2,U,V, r,g,b);
+		int g2 = bound((Y2 + v1[V] + u2[U]) >> 16);
-		*dst++ = 0xFF000000 | (r<<16) | (g<<8) | (b);
+		int r2 = bound((Y2 + v2[V]) >> 16);
+		*dst++ = 0xFF000000 | (r1<<16) | (g1<<8) | (b1);
+		*dst++ = 0xFF000000 | (r2<<16) | (g2<<8) | (b2);
 	}	
 }
 
@@ -69,18 +119,43 @@ void ConvertToXFB(u32 *dst, const u8* _pEFB, int width, int height)
 	const unsigned char *src = _pEFB;
 
 	u32 numBlocks = (width * height) / 2;
-
+#if 1
-	for (u32 i = 0; i < numBlocks; i++)
+	__m128i zero = _mm_setzero_si128();
+	for (int i = 0; i < numBlocks / 4; i++)
+	{
+		__m128i yuyv[4];
+		for (int j = 0; j < 4; j++) {
+			yuyv[j] = _mm_srai_epi32(
+				_mm_add_epi32(
+					_mm_add_epi32(
+						_mm_add_epi32(_r1[src[0]],	_mm_add_epi32(_g1[src[1]], _b1[src[2]])), _bias1),
+					_mm_add_epi32(
+						_mm_add_epi32(_r2[src[4]], _mm_add_epi32(_g2[src[5]], _b2[src[6]])), _bias2)
+					), 16);
+			src += 8;
+		}
+		__m128i four_dest = _mm_packus_epi16(_mm_packs_epi32(yuyv[0], yuyv[1]), _mm_packs_epi32(yuyv[2], yuyv[3]));
+		_mm_storeu_si128((__m128i *)dst, four_dest);
+		dst += 4;
+	}
+#else
+	for (int i = 0; i < numBlocks; i++)
 	{
 		int y1 = (((16843 * src[0]) + (33030 * src[1]) + (6423 * src[2])) >> 16) + 16;		
 		int u1 = ((-(9699 * src[0]) - (19071 * src[1]) + (28770 * src[2])) >> 16) + 128;
+
+		int v1 = (((28770 * src[0]) - (24117 * src[1]) - (4653 * src[2])) >> 16) + 128;
 		src += 4;
 
+		int u2 = ((-(9699 * src[0]) - (19071 * src[1]) + (28770 * src[2])) >> 16) + 128;
 		int y2 = (((16843 * src[0]) + (33030 * src[1]) + (6423 * src[2])) >> 16) + 16;
 		int v2 = (((28770 * src[0]) - (24117 * src[1]) - (4653 * src[2])) >> 16) + 128;
 		src += 4;
 
-		*dst++ = (v2 << 24) | (y2 << 16) | (u1 << 8) | (y1);
+		int u = bound_lut[(u1 + u2) / 2];
-	}	
+		int v = bound_lut[(v1 + v2) / 2];
+		*dst++ = (v << 24) | (y2 << 16) | (u << 8) | (y1);
+	}
+#endif
 }
 

Source/Core/VideoCommon/Src/XFBConvert.h

@@ -20,6 +20,8 @@
 
 #include "Common.h"
 
+void InitXFBConvTables();
+
 void ConvertFromXFB(u32 *dst, const u8* _pXFB, int width, int height);
 
 // converts 32-bit RGBA data to 16-bit 4:2:2 YUV data

Source/Core/VideoCommon/VideoCommon.vcproj

@@ -173,10 +173,15 @@
 			/>
 			<Tool
 				Name="VCCLCompilerTool"
+				Optimization="2"
+				EnableIntrinsicFunctions="true"
 				AdditionalIncludeDirectories="../../Core/Common/Src;../../PluginSpecs"
 				PreprocessorDefinitions="WIN32;NDEBUG;_LIB;_SECURE_SCL=0;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE"
+				StringPooling="true"
 				RuntimeLibrary="0"
 				BufferSecurityCheck="false"
+				EnableEnhancedInstructionSet="2"
+				FloatingPointModel="0"
 				UsePrecompiledHeader="0"
 				WarningLevel="3"
 				WarnAsError="false"
@@ -236,6 +241,10 @@
 			/>
 			<Tool
 				Name="VCCLCompilerTool"
+				Optimization="2"
+				EnableIntrinsicFunctions="true"
+				FavorSizeOrSpeed="1"
+				OmitFramePointers="false"
 				AdditionalIncludeDirectories="../../Core/Common/Src;../../PluginSpecs"
 				PreprocessorDefinitions="WIN32;NDEBUG;_LIB;_SECURE_SCL=0;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE"
 				RuntimeLibrary="0"
@@ -473,6 +482,14 @@
 		<File
 			RelativePath=".\Src\XFBConvert.cpp"
 			>
+			<FileConfiguration
+				Name="Release|Win32"
+				>
+				<Tool
+					Name="VCCLCompilerTool"
+					AssemblerOutput="4"
+				/>
+			</FileConfiguration>
 		</File>
 		<File
 			RelativePath=".\Src\XFBConvert.h"

Source/Plugins/Plugin_VideoDX9/Src/main.cpp

@@ -38,6 +38,7 @@
 #include "W32Util/Misc.h"
 #include "EmuWindow.h"
 #include "VideoState.h"
+#include "XFBConvert.h"
 
 #include "Utils.h"
 
@@ -125,7 +126,7 @@ bool Init()
 			return false;
 		}
 		InitLUTs();
-
+		InitXFBConvTables();
 	}
 	initCount++;
 

Source/Plugins/Plugin_VideoOGL/Src/OS/Win32.cpp

@@ -126,6 +126,8 @@ namespace EmuWindow
 			break;
 
 		case WM_CLOSE:
+			ExitProcess(0);
+
 			//Core::SetState(Core::CORE_UNINITIALIZED);
 			return 0;
 

Source/Plugins/Plugin_VideoOGL/Src/XFB.cpp

@@ -27,7 +27,9 @@
 
 enum {
 	XFB_WIDTH = 640,
-	XFB_HEIGHT = 480,
+	XFB_HEIGHT = 480, //480,
+	XFB_BUF_HEIGHT = 538, //480,
+	// TODO: figure out what to do with PAL
 };
 
 static GLuint xfb_texture;
@@ -39,8 +41,8 @@ static GLuint s_xfbRenderBuffer = 0;
 void XFB_Init()
 {
 	// used to render XFB
-	xfb_buffer = new u8[XFB_WIDTH * XFB_HEIGHT * 4];
+	xfb_buffer = new u8[XFB_WIDTH * XFB_BUF_HEIGHT * 4];
-	memset(xfb_buffer, 0, XFB_WIDTH * XFB_HEIGHT * 4);
+	memset(xfb_buffer, 0, XFB_WIDTH * XFB_BUF_HEIGHT * 4);
     glGenTextures(1, &xfb_texture);
 	glBindTexture(GL_TEXTURE_RECTANGLE_ARB, xfb_texture);
     glTexImage2D(GL_TEXTURE_RECTANGLE_ARB, 0, 4, XFB_WIDTH, XFB_HEIGHT, 0, GL_BGRA, GL_UNSIGNED_BYTE, xfb_buffer);

Source/Plugins/Plugin_VideoOGL/Src/main.cpp

@@ -37,6 +37,7 @@
 #include "PixelShaderManager.h"
 #include "VertexShaderManager.h"
 #include "XFB.h"
+#include "XFBConvert.h"
 
 #include "VideoState.h"
 #include "Debugger/Debugger.h" // for the CDebugger class
@@ -182,6 +183,7 @@ void Video_Initialize(SVideoInitialize* _pVideoInitialize)
     frameCount = 0;
     g_VideoInitialize = *_pVideoInitialize;
     InitLUTs();
+	InitXFBConvTables();
     g_Config.Load();
     
     if (!OpenGL_Create(g_VideoInitialize, 640, 480)) { //640x480 will be the default if all else fails//