VideoSoftware: Added Real XFB support.

Massivly increases the accuracy of VideoSoftware, if it's going to be slow, it
might as well be accurate.

 * Fixes flickering issues in games which double render to get Hardware AA, like the Peach's Castle tech demo.
 * Rouge Squadren 2's Cantina Intro doesn't flicker anymore, but it duplicates the top half of the screen onto the bottom.
 * Any games which need RealXFB support should now work in Video Software

Source/Plugins/Plugin_VideoSoftware/Src/EfbCopy.cpp

@@ -15,20 +15,26 @@
 #include "HW/Memmap.h"
 #include "Core.h"
 
+static const float s_gammaLUT[] =
+{
+	1.0f,
+	1.7f,
+	2.2f,
+	1.0f
+};
+
 namespace EfbCopy
 {
-	void CopyToXfb()
+	void CopyToXfb(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma)
 	{
-		GLInterface->Update(); // just updates the render window position and the backbuffer size	
-
 		if (!g_SWVideoConfig.bHwRasterizer)
 		{
-			// copy to open gl for rendering
+			INFO_LOG(VIDEO, "xfbaddr: %x, fbwidth: %i, fbheight: %i, source: (%i, %i, %i, %i), Gamma %f",
-			EfbInterface::UpdateColorTexture();
+					 xfbAddr, fbWidth, fbHeight, sourceRc.top, sourceRc.left, sourceRc.bottom, sourceRc.right, Gamma);
-			SWRenderer::DrawTexture(EfbInterface::efbColorTexture, EFB_WIDTH, EFB_HEIGHT);
+			EfbInterface::yuv422_packed* xfb_in_ram = (EfbInterface::yuv422_packed *) Memory::GetPointer(xfbAddr);
-		}
 
-		SWRenderer::SwapBuffer();
+			EfbInterface::CopyToXFB(xfb_in_ram, fbWidth, fbHeight, sourceRc, Gamma);
+		}
 	}
 
 	void CopyToRam()
@@ -47,8 +53,8 @@ namespace EfbCopy
 
 		int left   = bpmem.copyTexSrcXY.x;
 		int top    = bpmem.copyTexSrcXY.y;
-		int right  = left + bpmem.copyTexSrcWH.x;
+		int right  = left + bpmem.copyTexSrcWH.x + 1;
-		int bottom = top + bpmem.copyTexSrcWH.y;
+		int bottom = top + bpmem.copyTexSrcWH.y + 1;
 
 		for (u16 y = top; y <= bottom; y++)
 		{
@@ -62,21 +68,47 @@ namespace EfbCopy
 
 	void CopyEfb()
 	{
-		if (bpmem.triggerEFBCopy.copy_to_xfb)
+		EFBRectangle rc;
-			DebugUtil::OnFrameEnd();
+		rc.left = (int)bpmem.copyTexSrcXY.x;
+		rc.top = (int)bpmem.copyTexSrcXY.y;
+
+		// Here Width+1 like Height, otherwise some textures are corrupted already since the native resolution.
+		rc.right = (int)(bpmem.copyTexSrcXY.x + bpmem.copyTexSrcWH.x + 1);
+		rc.bottom = (int)(bpmem.copyTexSrcXY.y + bpmem.copyTexSrcWH.y + 1);
+
+		//if (bpmem.triggerEFBCopy.copy_to_xfb)
+		//    DebugUtil::OnFrameEnd(); // FIXME: not actually frame end
 
 		if (!g_bSkipCurrentFrame)
 		{
 			if (bpmem.triggerEFBCopy.copy_to_xfb)
 			{
-				CopyToXfb();
+				float yScale;
-				Core::Callback_VideoCopiedToXFB(true);
+				if (bpmem.triggerEFBCopy.scale_invert)
+					yScale = 256.0f / (float)bpmem.dispcopyyscale;
+				else
+					yScale = (float)bpmem.dispcopyyscale / 256.0f;
 
-				swstats.frameCount++;
+				float xfbLines = ((bpmem.copyTexSrcWH.y + 1.0f) * yScale);
+
+				if (yScale != 1.0)
+					WARN_LOG(VIDEO, "yScale of %f is currently unsupported", yScale);
+
+				if ((u32)xfbLines > MAX_XFB_HEIGHT)
+				{
+					INFO_LOG(VIDEO, "Tried to scale EFB to too many XFB lines (%f)", xfbLines);
+					xfbLines = MAX_XFB_HEIGHT;
+				}
+
+				CopyToXfb(bpmem.copyTexDest << 5,
+						  bpmem.copyMipMapStrideChannels << 4,
+						  (u32)xfbLines,
+						  rc,
+						  s_gammaLUT[bpmem.triggerEFBCopy.gamma]);
 			}
 			else
 			{
-				CopyToRam();
+				CopyToRam(); // FIXME: should use the rectangle we have already created above
 			}
 
 			if (bpmem.triggerEFBCopy.clear)
@@ -87,13 +119,5 @@ namespace EfbCopy
 					ClearEfb();
 			}
 		}
-		else
-		{
-			if (bpmem.triggerEFBCopy.copy_to_xfb)
-			{
-				// no frame rendered but tell that a frame has finished for frame skip counter
-				Core::Callback_VideoCopiedToXFB(false);
-			}
-		}
 	}
 }

Source/Plugins/Plugin_VideoSoftware/Src/EfbInterface.cpp

@@ -8,16 +8,13 @@
 #include "BPMemLoader.h"
 #include "LookUpTables.h"
 #include "SWPixelEngine.h"
+#include "HW/Memmap.h"
 
 
 u8 efb[EFB_WIDTH*EFB_HEIGHT*6];
 
-
 namespace EfbInterface
 {
-
-	u8 efbColorTexture[EFB_WIDTH*EFB_HEIGHT*4];
-
 	inline u32 GetColorOffset(u16 x, u16 y)
 	{
 		return (x + y * EFB_WIDTH) * 3;
@@ -31,7 +28,6 @@ namespace EfbInterface
 	void DoState(PointerWrap &p)
 	{
 		p.DoArray(efb, EFB_WIDTH*EFB_HEIGHT*6);
-		p.DoArray(efbColorTexture, EFB_WIDTH*EFB_HEIGHT*4);
 	}
 
 	void SetPixelAlphaOnly(u32 offset, u8 a)
@@ -469,6 +465,19 @@ namespace EfbInterface
 		GetPixelColor(offset, color);
 	}
 
+	// For internal used only, return a non-normalized value, which saves work later.
+	void GetColorYUV(u16 x, u16 y, yuv444 *out)
+	{
+		u8 color[4];
+		GetColor(x, y, color);
+
+		// GameCube/Wii uses the BT.601 standard algorithm for converting to YCbCr; see
+		// http://www.equasys.de/colorconversion.html#YCbCr-RGBColorFormatConversion
+		out->Y =  0.257f * color[RED_C] +  0.504f * color[GRN_C] +  0.098f * color[BLU_C];
+		out->U = -0.148f * color[RED_C] + -0.291f * color[GRN_C] +  0.439f * color[BLU_C];
+		out->V =  0.439f * color[RED_C] + -0.368f * color[GRN_C] + -0.071f * color[BLU_C];
+	}
+
 	u32 GetDepth(u16 x, u16 y)
 	{
 		u32 offset = GetDepthOffset(x, y);
@@ -482,22 +491,56 @@ namespace EfbInterface
 		return &efb[GetColorOffset(x, y)];
 	}
 
-	void UpdateColorTexture()
+	void CopyToXFB(yuv422_packed* xfb_in_ram, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma) {
-	{
+		// FIXME: We should do Gamma correction
-		u32 color;
-		u8* colorPtr = (u8*)&color;
-		u32* texturePtr = (u32*)efbColorTexture;
-		u32 textureAddress = 0;
-		u32 efbOffset = 0;
 
-		for (u16 y = 0; y < EFB_HEIGHT; y++)
+		if (!xfb_in_ram)
 		{
-			for (u16 x = 0; x < EFB_WIDTH; x++)
+			WARN_LOG(VIDEO, "Tried to copy to invalid XFB address");
+			return;
+		}
+
+		int left = sourceRc.left;
+		int right = sourceRc.right;
+
+		// this assumes copies will always start on an even (YU) pixel and the
+		// copy always has an even width, which might not be true.
+		if (left & 1 || right & 1) {
+			WARN_LOG(VIDEO, "Trying to copy XFB to from unaligned EFB source");
+			// this will show up as wrongly encoded
+		}
+
+		// Scanline buffer, leave room for borders
+		yuv444 scanline[640+2];
+
+		// our internal yuv444 type is not normalized, so black is {0, 0, 0} instead of {16, 128, 128}
+		scanline[0] = {0, 0, 0}; // black border at start
+		scanline[right+1] = {0, 0, 0}; // black border at end
+
+		for (u16 y = sourceRc.top; y < sourceRc.bottom; y++)
+		{
+			// Get a scanline of YUV pixels in 4:4:4 format
+
+			for (int i = 1, x = left; x < right; i++, x++)
 			{
-				GetPixelColor(efbOffset, colorPtr);
+				GetColorYUV(x, y, &scanline[i]);
-				efbOffset += 3;
-				texturePtr[textureAddress++] = Common::swap32(color);  // ABGR->RGBA
 			}
+
+			// And Downsample them to 4:2:2
+			for (int i = 1, x = left; x < right; i+=2, x+=2)
+			{
+				// YU pixel
+				xfb_in_ram[x].Y = scanline[i].Y;
+				// U[i] = 1/4 * U[i-1] + 1/2 * U[i] + 1/4 U[i+1]
+				// we add in 10 bit space so it will round more accurately
+				xfb_in_ram[x].UV = 128 + ((scanline[i-1].U + (scanline[i].U << 1) + scanline[i+1].U) >> 2);
+
+				// YV pixel
+				xfb_in_ram[x+1].Y = scanline[i+1].Y;
+				// V[i] = 1/4 * V[i-1] + 1/2 * V[i] + 1/4 V[i+1]
+				xfb_in_ram[x+1].UV = 128 + ((scanline[i].V + (scanline[i+1].V << 1) + scanline[i+2].V) >> 2);
+			}
+			xfb_in_ram += 640;
 		}
 	}
 

Source/Plugins/Plugin_VideoSoftware/Src/EfbInterface.h

@@ -11,9 +11,21 @@ namespace EfbInterface
 {
 	const int DEPTH_BUFFER_START = EFB_WIDTH * EFB_HEIGHT * 3;
 
+	// color order is ABGR in order to emulate RGBA on little-endian hardware
 	enum { ALP_C, BLU_C, GRN_C, RED_C };
 
-	// color order is ABGR in order to emulate RGBA on little-endian hardware
+	// packed so the compiler doesn't mess with alignment
+	typedef struct __attribute__ ((packed)) {
+		u8 Y;
+		u8 UV;
+	} yuv422_packed;
+
+	// But this one is only used internally, so we can let the compiler pack it however it likes.
+	typedef struct __attribute__ ((aligned (4))){
+		u8 Y;
+		s8 U;
+		s8 V;
+	} yuv444;
 
 	// does full blending of an incoming pixel
 	void BlendTev(u16 x, u16 y, u8 *color);
@@ -28,12 +40,13 @@ namespace EfbInterface
 	void SetDepth(u16 x, u16 y, u32 depth);
 
 	void GetColor(u16 x, u16 y, u8 *color);
+	void GetColorYUV(u16 x, u16 y, yuv444 *color);
 	u32 GetDepth(u16 x, u16 y);
 
 	u8* GetPixelPointer(u16 x, u16 y, bool depth);
 
-	void UpdateColorTexture();
+	void CopyToXFB(yuv422_packed* xfb_in_ram, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma);
-	extern u8 efbColorTexture[EFB_WIDTH*EFB_HEIGHT*4]; // RGBA format
+
 	void DoState(PointerWrap &p);
 }
 

Source/Plugins/Plugin_VideoSoftware/Src/SWRenderer.cpp

@@ -5,10 +5,12 @@
 #include "Common.h"
 #include <math.h>
 
+#include "Core.h"
 #include "../../Plugin_VideoOGL/Src/GLUtil.h"
 #include "RasterFont.h"
 #include "SWRenderer.h"
 #include "SWStatistics.h"
+#include "SWCommandProcessor.h"
 
 #include "OnScreenDisplay.h"
 
@@ -18,6 +20,9 @@ static GLint attr_pos = -1, attr_tex = -1;
 static GLint uni_tex = -1;
 static GLuint program;
 
+static u8 s_xfbColorTexture[2][EFB_WIDTH*EFB_HEIGHT*4];
+static int s_currentColorTexture = 0;
+
 // Rasterfont isn't compatible with GLES
 // degasus: I think it does, but I can't test it
 #ifndef USE_GLES
@@ -26,6 +31,7 @@ RasterFont* s_pfont = NULL;
 
 void SWRenderer::Init()
 {
+    GLInterface->SetBackBufferDimensions(EFB_WIDTH, EFB_HEIGHT);
 }
 
 void SWRenderer::Shutdown()
@@ -68,6 +74,9 @@ void CreateShaders()
 
 void SWRenderer::Prepare()
 {
+	memset(s_xfbColorTexture, 0, sizeof(s_xfbColorTexture));
+	s_currentColorTexture = 0;
+
 	glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
 	glPixelStorei(GL_UNPACK_ALIGNMENT, 4);  // 4-byte pixel alignment
 	glGenTextures(1, &s_RenderTarget);
@@ -162,11 +171,59 @@ void DrawButton(GLuint tex, float *coords)
         glBindTexture(TEX2D, 0); 
 }
 #endif
+
+void SWRenderer::UpdateColorTexture(EfbInterface::yuv422_packed *xfb)
+{
+	u32 offset = 0;
+	u8 *TexturePointer = s_xfbColorTexture[!s_currentColorTexture];
+
+	for (u16 y = 0; y < EFB_HEIGHT; y++)
+	{
+		for (u16 x = 0; x < EFB_WIDTH; x+=2)
+		{
+			// We do this one color sample (aka 2 RGB pixles) at a time
+			int Y1 = xfb[x].Y - 16;
+			int Y2 = xfb[x+1].Y - 16;
+			int U  = int(xfb[x].UV) - 128;
+			int V  = int(xfb[x+1].UV) - 128;
+
+			// We do the inverse BT.601 conversion for YCbCr to RGB
+			// http://www.equasys.de/colorconversion.html#YCbCr-RGBColorFormatConversion
+			TexturePointer[offset++] = min(255.0f, max(0.0f, 1.164f * Y1              + 1.596f * V));
+			TexturePointer[offset++] = min(255.0f, max(0.0f, 1.164f * Y1 - 0.392f * U - 0.813f * V));
+			TexturePointer[offset++] = min(255.0f, max(0.0f, 1.164f * Y1 + 2.017f * U             ));
+			TexturePointer[offset++] = 255;
+
+			TexturePointer[offset++] = min(255.0f, max(0.0f, 1.164f * Y2              + 1.596f * V));
+			TexturePointer[offset++] = min(255.0f, max(0.0f, 1.164f * Y2 - 0.392f * U - 0.813f * V));
+			TexturePointer[offset++] = min(255.0f, max(0.0f, 1.164f * Y2 + 2.017f * U             ));
+			TexturePointer[offset++] = 255;
+		}
+		xfb += EFB_WIDTH;
+	}
+	s_currentColorTexture = !s_currentColorTexture;
+}
+
+// Called on the GPU thread
+void SWRenderer::Swap(u32 fbWidth, u32 fbHeight)
+{
+	GLInterface->Update(); // just updates the render window position and the backbuffer size
+	if (!g_SWVideoConfig.bHwRasterizer)
+		SWRenderer::DrawTexture(s_xfbColorTexture[s_currentColorTexture], fbWidth, fbHeight);
+
+	swstats.frameCount++;
+	SWRenderer::SwapBuffer();
+	Core::Callback_VideoCopiedToXFB(true); // FIXME: should this function be called FrameRendered?
+}
+
 void SWRenderer::DrawTexture(u8 *texture, int width, int height)
 {
+	// FIXME: This should add black bars when the game has set the VI to render less than the full xfb.
+
 	GLsizei glWidth = (GLsizei)GLInterface->GetBackBufferWidth();
 	GLsizei glHeight = (GLsizei)GLInterface->GetBackBufferHeight();
 
+
 	// Update GLViewPort
 	glViewport(0, 0, glWidth, glHeight);
 	glScissor(0, 0, glWidth, glHeight);

Source/Plugins/Plugin_VideoSoftware/Src/SWRenderer.h

@@ -6,6 +6,7 @@
 #define _RENDERER_H_
 
 #include "CommonTypes.h"
+#include "EfbInterface.h"
 
 namespace SWRenderer
 {
@@ -16,8 +17,10 @@ namespace SWRenderer
 	void RenderText(const char* pstr, int left, int top, u32 color);
 	void DrawDebugText();
 
+	void UpdateColorTexture(EfbInterface::yuv422_packed *xfb);
 	void DrawTexture(u8 *texture, int width, int height);
 
+	void Swap(u32 fbWidth, u32 fbHeight);
 	void SwapBuffer();
 }
 

Source/Plugins/Plugin_VideoSoftware/Src/SWmain.cpp

@@ -9,7 +9,7 @@
 #include "VideoConfigDialog.h"
 #endif // HAVE_WX
 
-
+#include "Atomic.h"
 #include "SWCommandProcessor.h"
 #include "OpcodeDecoder.h"
 #include "SWVideoConfig.h"
@@ -29,10 +29,22 @@
 #include "OpcodeDecoder.h"
 #include "SWVertexLoader.h"
 #include "SWStatistics.h"
+#include "HW/VideoInterface.h"
+#include "HW/Memmap.h"
 
 #include "OnScreenDisplay.h"
 #define VSYNC_ENABLED 0
 
+static volatile u32 s_swapRequested = false;
+
+static volatile struct
+{
+    u32 xfbAddr;
+    FieldType field;
+    u32 fbWidth;
+    u32 fbHeight;
+} s_beginFieldArgs;
+
 namespace SW
 {
 
@@ -191,12 +203,51 @@ void VideoSoftware::Video_Prepare()
 
 // Run from the CPU thread (from VideoInterface.cpp)
 void VideoSoftware::Video_BeginField(u32 xfbAddr, FieldType field, u32 fbWidth, u32 fbHeight)
-{	
+{
+	s_beginFieldArgs.xfbAddr = xfbAddr;
+	s_beginFieldArgs.field = field;
+	s_beginFieldArgs.fbWidth = fbWidth;
+	s_beginFieldArgs.fbHeight = fbHeight;
 }
 
 // Run from the CPU thread (from VideoInterface.cpp)
 void VideoSoftware::Video_EndField()
 {
+	// Techincally the XFB is continually rendered out scanline by scanline between
+	// BeginField and EndFeild, We could possibly get away with copying out the whole thing
+	// at BeginField for less lag, but for the safest emulation we run it here.
+
+	if (g_bSkipCurrentFrame || s_beginFieldArgs.xfbAddr == 0 ) {
+		swstats.frameCount++;
+		swstats.ResetFrame();
+		Core::Callback_VideoCopiedToXFB(false);
+		return;
+	}
+	if (!g_SWVideoConfig.bHwRasterizer) {
+
+		// Force Progressive
+		u32 xfbAddr = VideoInterface::GetXFBAddressTop();
+
+		// All drivers make an assumption that the two fields are interleaved in the framebuffer
+		// Give a warning if this isn't true.
+		if (xfbAddr + 1280 != VideoInterface::GetXFBAddressBottom()) {
+			WARN_LOG(VIDEO, "Feilds are not interleaved in XFB as expected.");
+		}
+
+		EfbInterface::yuv422_packed *xfb = (EfbInterface::yuv422_packed *) Memory::GetPointer(xfbAddr);
+
+		SWRenderer::UpdateColorTexture(xfb);
+	}
+
+	// Idealy we would just move all the opengl contex stuff to the CPU thread, but this gets
+	// messy when the Hardware Rasterizer is enabled.
+	// And Neobrain loves his Hardware Rasterizer
+
+	// If we are runing dual core, Signal the GPU thread about the new colour texture.
+	if (SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread)
+		Common::AtomicStoreRelease(s_swapRequested, true);
+	else
+		SWRenderer::Swap(s_beginFieldArgs.fbWidth, s_beginFieldArgs.fbHeight);
 }
 
 u32 VideoSoftware::Video_AccessEFB(EFBAccessType type, u32 x, u32 y, u32 InputData)
@@ -242,6 +293,16 @@ bool VideoSoftware::Video_Screenshot(const char *_szFilename)
 	return false;
 }
 
+// Run from the graphics thread
+static void VideoFifo_CheckSwapRequest()
+{
+	if (Common::AtomicLoadAcquire(s_swapRequested))
+	{
+		SWRenderer::Swap(s_beginFieldArgs.fbWidth, s_beginFieldArgs.fbHeight);
+		Common::AtomicStoreRelease(s_swapRequested, false);
+	}
+}
+
 // -------------------------------
 // Enter and exit the video loop
 // -------------------------------
@@ -252,6 +313,7 @@ void VideoSoftware::Video_EnterLoop()
 
 	while (fifoStateRun)
 	{
+		VideoFifo_CheckSwapRequest();
 		g_video_backend->PeekMessages();
 
 		if (!SWCommandProcessor::RunBuffer())
@@ -262,11 +324,12 @@ void VideoSoftware::Video_EnterLoop()
 		while (!emuRunningState && fifoStateRun)
 		{
 			g_video_backend->PeekMessages();
+			VideoFifo_CheckSwapRequest();
 			m_csSWVidOccupied.unlock();
 			Common::SleepCurrentThread(1);
 			m_csSWVidOccupied.lock();
 		}
-	}	
+	}
 }
 
 void VideoSoftware::Video_ExitLoop()